arcjet

PyPI badge

Arcjet is the runtime security platform that ships in your AI code. Detect prompt injection, authorize agent tool calls, redact sensitive data, and block bots and abuse. Real-time security building blocks you call inside your app, before an action happens.

This is the Python SDK for Arcjet — use arcjet / arcjet_sync for request protection (FastAPI, Flask, Django route handlers) and arcjet.guard for guard protection (AI agent tool calls, MCP servers, background jobs).

Why Arcjet?

Your app's AI features and agents take real actions, calling tools, reading data, hitting APIs. Arcjet runs inside that code and lets you enforce security on each action in real time, then audit what happened.

Getting started

Install the Arcjet CLI

The CLI is used to log in, manage site keys, and install protection skills.

Homebrew (macOS and Linux):

brew install arcjet/tap/arcjet

npx (Node.js) — run any command without installing:

npx @arcjet/cli <command>

Or download a binary for macOS (Apple Silicon, Intel), Linux (x86_64, arm64), and Windows (x86_64, arm64).

Examples below use the arcjet binary. If you installed via npx, replace arcjet with npx @arcjet/cli.

Quick setup with an AI agent

Log in with the CLI:
```
arcjet auth login
```
Install the Arcjet skill:
```
npx skills add arcjet/skills
```
Tell your agent what to protect — it handles the rest.

Manual setup

Log in with the CLI (or at app.arcjet.com):
```
arcjet auth login
```
pip install arcjet (or uv add arcjet)
Get your site key:
```
arcjet sites get-key
```
Or copy it from the Arcjet dashboard.
Set ARCJET_KEY=ajkey_yourkey in .env
Protect a route — see the AI protection example or individual feature examples below.

Get help

Join our Discord server or reach out for support.

Documentation — full reference and guides
Examples — FastAPI and Flask example apps, including LangChain integration
Blueprints — recipes for common security patterns

Quick start

Note: Examples below use FastAPI (async). For Flask and other sync frameworks, use arcjet_sync instead of arcjet. The API is identical — see Async vs. sync client.

Protect an AI chat endpoint with prompt injection detection, token budget rate limiting, and bot protection:

# main.py
import os
from fastapi import FastAPI, Request
from fastapi.responses import JSONResponse
from pydantic import BaseModel

from arcjet import (
    arcjet,  # async client — use arcjet_sync for Flask and other sync frameworks
    detect_bot,
    detect_prompt_injection,
    detect_sensitive_info,
    shield,
    token_bucket,
    Mode,
    SensitiveInfoEntityType,
)

app = FastAPI()

arcjet_key = os.getenv("ARCJET_KEY")
if not arcjet_key:
    raise RuntimeError(
        "ARCJET_KEY is required. Get one with: arcjet sites get-key"
        " or from https://app.arcjet.com"
    )

# Create a single Arcjet instance and reuse it across requests.
# Use arcjet_sync instead if you are using Flask or another sync framework.
aj = arcjet(
    key=arcjet_key,
    rules=[
        # Detect and block prompt injection attacks in user messages
        detect_prompt_injection(mode=Mode.LIVE),
        # Block sensitive data (e.g. credit cards, PII) from reaching your LLM
        detect_sensitive_info(
            mode=Mode.LIVE,
            deny=[
                SensitiveInfoEntityType.CREDIT_CARD_NUMBER,
                SensitiveInfoEntityType.EMAIL,
                SensitiveInfoEntityType.PHONE_NUMBER,
            ],
        ),
        # Rate limit by token budget — refill 100 tokens every 60 seconds
        token_bucket(
            characteristics=["userId"],
            mode=Mode.LIVE,
            refill_rate=100,
            interval=60,
            capacity=1000,
        ),
        # Block automated clients and scrapers from your AI endpoints
        detect_bot(
            mode=Mode.LIVE,
            allow=[],  # empty = block all bots
        ),
        # Protect against common web attacks (SQLi, XSS, etc.)
        shield(mode=Mode.LIVE),
    ],
)


class ChatRequest(BaseModel):
    message: str


@app.post("/chat")
async def chat(request: Request, body: ChatRequest):
    userId = "user_123"  # replace with real user ID from session

    decision = await aj.protect(
        request,
        requested=5,  # tokens consumed per request
        characteristics={"userId": userId},
        detect_prompt_injection_message=body.message,  # scan for prompt injection
        sensitive_info_value=body.message,  # scan for PII
    )

    if decision.is_denied():
        status = 429 if decision.reason_v2.type == "RATE_LIMIT" else 403
        return JSONResponse({"error": "Denied"}, status_code=status)

    # Safe to pass body.message to your LLM
    return {"reply": "..."}

Features

Feature	Request (`arcjet`)	Guard (`arcjet.guard`)
Rate Limiting	✅	✅
Prompt Injection Detection	✅	✅
Content Moderation	—	✅
Sensitive Information Detection	✅	✅
Bot Protection	✅	—
Shield WAF	✅	—
Email Validation	✅	—
Request Filters	✅	—
IP Analysis	✅	—
Custom Rules	—	✅
Capture (visibility events)	—	✅

🔒 Prompt Injection Detection — detect and block prompt injection attacks before they reach your LLM.
🚫 Content Moderation — detect harmful content in Guard tool-call and job inputs.
🤖 Bot Protection — stop scrapers, credential stuffers, and AI crawlers from abusing your endpoints.
🛑 Rate Limiting — token bucket, fixed window, and sliding window algorithms; model AI token budgets per user.
🕵️ Sensitive Information Detection — block PII, credit cards, and custom patterns from entering your AI pipeline.
🛡️ Shield WAF — protect against SQL injection, XSS, and other common web attacks.
📧 Email Validation — block disposable, invalid, and undeliverable addresses at signup.
📝 Signup Form Protection — combines bot protection, email validation, and rate limiting to protect your signup forms.
🎯 Request Filters — expression-based rules on IP, path, headers, and custom fields.
🌐 IP Analysis — geolocation, ASN, VPN, proxy, Tor, and hosting detection included with every request.
🧩 Arcjet Guard — lower-level API for AI agent tool calls and background tasks where there is no HTTP request.

Which features do I need?

If your app has...	Recommended features
LLM / AI chat endpoints	Prompt injection + sensitive info + token bucket rate limit + bot protection + shield
AI agent tool calls	Arcjet Guard — rate limiting + prompt injection + content moderation + sensitive info + custom rules
Public API	Rate limiting + bot protection + shield
Signup / login forms	Email validation + bot protection + rate limiting (or signup protection)
Internal / admin routes	Shield + request filters (country, VPN/proxy blocking)
Any web application	Shield + bot protection (good baseline for all apps)

All features can be combined in a single Arcjet instance. Rules are evaluated together — if any rule denies the request, decision.is_denied() returns True. Use Mode.DRY_RUN on individual rules to test them before enforcing.

Installation

Install from PyPI with uv:

# With a uv project
uv add arcjet

# With an existing pip managed project
uv pip install arcjet

Or with pip:

pip install arcjet

Prefer a glibc Linux image (Debian/Ubuntu slim) in containers. Alpine/musl is not a supported install target — see Compatibility.

Prompt injection detection

Detect and block prompt injection attacks — attempts by users to hijack your LLM's behavior through crafted input — before they reach your model.

FastAPI

from arcjet import arcjet, detect_prompt_injection, Mode

aj = arcjet(
    key=arcjet_key,
    rules=[
        detect_prompt_injection(mode=Mode.LIVE),
    ],
)


@app.post("/chat")
async def chat(request: Request, body: ChatRequest):
    decision = await aj.protect(
        request,
        detect_prompt_injection_message=body.message,
    )

    if decision.is_denied():
        return JSONResponse({"error": "Prompt injection detected"}, status_code=403)

    # safe to pass body.message to your LLM

Flask

from arcjet import arcjet_sync, detect_prompt_injection, Mode

aj = arcjet_sync(
    key=arcjet_key,
    rules=[
        detect_prompt_injection(mode=Mode.LIVE),
    ],
)


@app.route("/chat", methods=["POST"])
def chat():
    body = request.get_json()
    decision = aj.protect(request, detect_prompt_injection_message=body["message"])

    if decision.is_denied():
        return jsonify(error="Prompt injection detected"), 403

    # safe to pass body["message"] to your LLM

See the Prompt Injection docs for more details.

Bot protection

Manage traffic from automated clients. Block scrapers, credential stuffers, and AI crawlers, while allowing legitimate bots like search engines and monitors.

FastAPI

from arcjet import arcjet, detect_bot, Mode, BotCategory

aj = arcjet(
    key=arcjet_key,
    rules=[
        detect_bot(
            mode=Mode.LIVE,
            allow=[
                BotCategory.SEARCH_ENGINE,  # Google, Bing, etc.
                # BotCategory.MONITOR,      # Uptime monitoring
                # BotCategory.PREVIEW,      # Link previews (Slack, Discord)
                # "OPENAI_CRAWLER_SEARCH",  # Allow OpenAI crawler
            ],
        ),
    ],
)


@app.get("/")
async def index(request: Request):
    decision = await aj.protect(request)

    if decision.is_denied():
        return JSONResponse({"error": "Bot detected"}, status_code=403)

    return {"message": "Hello world"}

Flask

from arcjet import arcjet_sync, detect_bot, is_spoofed_bot, Mode, BotCategory

aj = arcjet_sync(
    key=arcjet_key,
    rules=[
        detect_bot(mode=Mode.LIVE, allow=[BotCategory.SEARCH_ENGINE]),
    ],
)


@app.route("/")
def index():
    decision = aj.protect(request)

    if decision.is_denied():
        return jsonify(error="Bot detected"), 403

    if any(is_spoofed_bot(r) for r in decision.results):
        return jsonify(error="Spoofed bot"), 403

    return jsonify(message="Hello world")

Bot categories

Configure rules using categories or specific bot identifiers:

detect_bot(
    mode=Mode.LIVE,
    allow=[
        BotCategory.SEARCH_ENGINE,
        "OPENAI_CRAWLER_SEARCH",
    ],
)

Available categories: ACADEMIC, ADVERTISING, AI, AMAZON, APPLE, ARCHIVE, BOTNET, FEEDFETCHER, GOOGLE, META, MICROSOFT, MONITOR, OPTIMIZER, PREVIEW, PROGRAMMATIC, SEARCH_ENGINE, SLACK, SOCIAL, TOOL, UNKNOWN, VERCEL, WEBHOOK, YAHOO. Use BotCategory.<NAME> in Python or pass the string directly. You can also allow or deny specific bots by name.

If you specify an allow list, all other bots are denied. An empty allow list blocks all bots. The reverse applies for deny lists.

Verified vs. spoofed bots

Bots claiming to be well-known crawlers (e.g. Googlebot) are verified against their known IP ranges. Use is_spoofed_bot() to check:

from arcjet import is_spoofed_bot

if any(is_spoofed_bot(r) for r in decision.results):
    return jsonify(error="Spoofed bot"), 403

See the Bot Protection docs for more details.

Rate limiting

Limit request rates per IP, user, or any custom characteristic. Arcjet supports token bucket, fixed window, and sliding window algorithms. Token buckets are ideal for controlling AI token budgets — set capacity to the max tokens a user can spend, refill_rate to how many tokens are restored per interval, and deduct tokens per request via requested in protect(). The interval accepts seconds as a number. Use characteristics to track limits per user instead of per IP.

Token bucket (recommended for AI)

Rate limits track by IP address by default. To track per user, declare the key name in characteristics on the rule, then pass the actual value in protect():

from arcjet import arcjet, token_bucket, Mode

aj = arcjet(
    key=arcjet_key,
    rules=[
        token_bucket(
            characteristics=["userId"],  # or ["ip.src"] for IP-based
            mode=Mode.LIVE,
            refill_rate=100,  # tokens added per interval
            interval=60,  # interval in seconds
            capacity=1000,  # maximum tokens per bucket
        ),
    ],
)


@app.post("/chat")
async def chat(request: Request):
    decision = await aj.protect(
        request,
        requested=5,  # tokens consumed by this request
        characteristics={"userId": "user_123"},
    )

    if decision.is_denied():
        return JSONResponse({"error": "Rate limited"}, status_code=429)

Fixed window

from arcjet import arcjet, fixed_window, Mode

aj = arcjet(
    key=arcjet_key,
    rules=[
        fixed_window(mode=Mode.LIVE, window=60, max=100),
    ],
)

Sliding window

from arcjet import arcjet, sliding_window, Mode

aj = arcjet(
    key=arcjet_key,
    rules=[
        sliding_window(mode=Mode.LIVE, interval=60, max=100),
    ],
)

See the Rate Limiting docs for more details.

Sensitive information detection

Detect and block PII in request content before it reaches your LLM or data store. The default (local WebAssembly) backend detects EMAIL, PHONE_NUMBER, IP_ADDRESS, and CREDIT_CARD_NUMBER. You can provide a custom detect callback for additional patterns, or the optional on-device Rampart backend (see below) for names, addresses, and government/financial identifiers.

from arcjet import arcjet, detect_sensitive_info, SensitiveInfoEntityType, Mode

aj = arcjet(
    key=arcjet_key,
    rules=[
        detect_sensitive_info(
            mode=Mode.LIVE,
            deny=[
                SensitiveInfoEntityType.EMAIL,
                SensitiveInfoEntityType.CREDIT_CARD_NUMBER,
            ],
        ),
    ],
)

# Pass the content to scan with each protect() call
decision = await aj.protect(request, sensitive_info_value="User input to scan")

You can supplement built-in detectors with a custom detect callback:

def my_detect(tokens: list[str]) -> list[str | None]:
    return ["CUSTOM_PII" if "secret" in t.lower() else None for t in tokens]


rules = [
    detect_sensitive_info(
        mode=Mode.LIVE,
        deny=["CUSTOM_PII"],
        detect=my_detect,
    ),
]

On-device Rampart backend (more entity types)

The default backend detects the four types above. To detect names, addresses, and government/financial identifiers, install the optional arcjet[sensitive-info-rampart] extra and pass its backend to the rule. It runs the on-device Rampart NER model entirely locally, so no data leaves your environment:

pip install "arcjet[sensitive-info-rampart]"

from arcjet import arcjet, detect_sensitive_info, Mode
from arcjet_sensitive_info_rampart import rampart

aj = arcjet(
    key=arcjet_key,
    rules=[
        detect_sensitive_info(
            mode=Mode.LIVE,
            deny=["EMAIL", "GIVEN_NAME", "SURNAME", "STREET_NAME", "SSN"],
            backend=rampart(),
        ),
    ],
)

The backend adds these entity types: GIVEN_NAME, SURNAME, SSN, URL, TAX_ID, BANK_ACCOUNT, ROUTING_NUMBER, GOVERNMENT_ID, PASSPORT, DRIVERS_LICENSE, BUILDING_NUMBER, STREET_NAME, SECONDARY_ADDRESS, CITY, STATE, ZIP_CODE. Listing one of these without a supporting backend (or a custom detect function) raises, since the default engine can never match it. The bundled model loads once on first use and is reused for every request. See the arcjet-sensitive-info-rampart README and the examples/fastapi-rampart example.

See the Sensitive Information docs for more details.

Shield WAF

Protect against common web attacks including SQL injection, XSS, path traversal, and other OWASP Top 10 threats. No additional configuration needed — Shield analyzes request patterns automatically.

from arcjet import arcjet, shield, Mode

aj = arcjet(
    key=arcjet_key,
    rules=[
        shield(mode=Mode.LIVE),
    ],
)

See the Shield docs for more details.

Email validation

Prevent users from signing up with disposable, invalid, or undeliverable email addresses. Deny types: DISPOSABLE, FREE, INVALID, NO_MX_RECORDS, NO_GRAVATAR.

from arcjet import arcjet, validate_email, EmailType, Mode

aj = arcjet(
    key=arcjet_key,
    rules=[
        validate_email(
            mode=Mode.LIVE,
            deny=[
                EmailType.DISPOSABLE,
                EmailType.INVALID,
                EmailType.NO_MX_RECORDS,
            ],
        ),
    ],
)

# Pass the email with each protect() call
decision = await aj.protect(request, email="user@example.com")

See the Email Validation docs for more details.

Request filters

Filter requests using expression-based rules against request properties (IP address, headers, path, HTTP method, and custom local fields).

Block by country

Restrict access to specific countries — useful for licensing, compliance, or regional rollouts. The allow list denies all countries not listed:

from arcjet import arcjet, filter_request, Mode

aj = arcjet(
    key=arcjet_key,
    rules=[
        # Allow only US traffic — all other countries are denied
        filter_request(
            mode=Mode.LIVE,
            allow=['ip.src.country eq "US"'],
        ),
    ],
)


@app.get("/")
async def index(request: Request):
    decision = await aj.protect(request)

    if decision.is_denied():
        return JSONResponse(
            {"error": "Access restricted in your region"}, status_code=403
        )

To restrict to a specific state or province, combine country and region:

filter_request(
    mode=Mode.LIVE,
    # Allow only California — useful for state-level compliance e.g. CCPA testing
    allow=['ip.src.country eq "US" and ip.src.region eq "California"'],
)

Block VPN and proxy traffic

Prevent anonymized traffic from accessing sensitive endpoints — useful for fraud prevention, enforcing geo-restrictions, and reducing abuse:

from arcjet import arcjet, filter_request, Mode

aj = arcjet(
    key=arcjet_key,
    rules=[
        filter_request(
            mode=Mode.LIVE,
            deny=[
                "ip.src.vpn",  # VPN services
                "ip.src.proxy",  # Open proxies
                "ip.src.tor",  # Tor exit nodes
            ],
        ),
    ],
)

For cases where you want to allow some anonymized traffic (e.g. Apple Private Relay) but still log or handle it differently, use decision.ip helpers after calling protect():

decision = await aj.protect(request)

if decision.ip.is_vpn() or decision.ip.is_tor():
    return JSONResponse({"error": "VPN traffic not allowed"}, status_code=403)

ip = decision.ip_details
if ip and ip.is_relay:
    # Privacy relay (e.g. Apple Private Relay) — lower risk than a VPN
    pass  # allow through with custom handling

Custom local fields

Pass arbitrary values from your application for use in filter expressions:

decision = await aj.protect(
    request,
    filter_local={"userId": current_user.id, "plan": current_user.plan},
)

These are then available as local.userId and local.plan in expressions:

filter_request(
    mode=Mode.LIVE,
    deny=['local.plan eq "free" and ip.src.country ne "US"'],
)

See the Request Filters docs, IP Geolocation blueprint, and VPN/Proxy Detection blueprint for more details.

IP analysis

Arcjet returns IP metadata with every decision — no extra API calls needed.

# High-level helpers
if decision.ip.is_hosting():
    # likely a cloud/hosting provider — often suspicious for bots
    return JSONResponse({"error": "Hosting IP blocked"}, status_code=403)

if decision.ip.is_vpn() or decision.ip.is_proxy() or decision.ip.is_tor():
    # apply your policy for anonymized traffic
    pass

# Typed field access
ip = decision.ip_details
if ip:
    print(ip.city, ip.country_name)  # geolocation
    print(ip.asn, ip.asn_name)  # ASN / network
    print(ip.is_vpn, ip.is_hosting)  # reputation
    if ip.threat:  # optional threat intelligence
        threat = ip.threat
        print(threat.risk_level, threat.confidence, threat.reputation)
        print(threat.is_safe, threat.network_types, threat.activities)
        print(threat.entities, threat.entity_name, threat.service)

Available fields include geolocation (latitude, longitude, city, region, country, continent), network (asn, asn_name, asn_domain, asn_type, asn_country), and reputation (is_vpn, is_proxy, is_tor, is_hosting, is_relay). Threat intelligence provides risk_level, confidence, reputation, is_safe, network_types, activities, entities, entity_name, and service.

decision.ip_details and its threat field are optional because metadata or threat intelligence may not be available for every IP.

LangChain example

Arcjet works with any Python code, including LangChain agents and chains. In this example, we protect a LangChain agent's chat endpoint with Arcjet to prevent prompt injection, block bots, prevent sensitive data leakage, and enforce token budgets before invoking the agent.

FastAPI + LangChain

from arcjet import (
    arcjet,
    detect_bot,
    detect_prompt_injection,
    detect_sensitive_info,
    token_bucket,
    Mode,
    SensitiveInfoEntityType,
)

aj = arcjet(
    key=arcjet_key,
    rules=[
        detect_prompt_injection(mode=Mode.LIVE),
        detect_sensitive_info(
            mode=Mode.LIVE,
            deny=[
                SensitiveInfoEntityType.EMAIL,
                SensitiveInfoEntityType.CREDIT_CARD_NUMBER,
                SensitiveInfoEntityType.PHONE_NUMBER,
            ],
        ),
        detect_bot(mode=Mode.LIVE, allow=["CURL"]),
        token_bucket(
            characteristics=["userId"],
            mode=Mode.LIVE,
            refill_rate=5,
            interval=10,
            capacity=10,
        ),
    ],
)


@app.post("/chat")
async def chat(request: Request, body: ChatRequest):
    decision = await aj.protect(
        request,
        requested=5,
        characteristics={"userId": "user_123"},
        detect_prompt_injection_message=body.message,  # scan for prompt injection
        sensitive_info_value=body.message,  # scan for PII before sending to LLM
    )

    if decision.is_denied():
        status = 429 if decision.reason_v2.type == "RATE_LIMIT" else 403
        return JSONResponse({"error": "Denied"}, status_code=status)

    reply = await chain.ainvoke({"message": body.message})
    return {"reply": reply}

Flask + LangChain

from arcjet import (
    arcjet_sync,
    detect_bot,
    detect_prompt_injection,
    detect_sensitive_info,
    token_bucket,
    Mode,
    SensitiveInfoEntityType,
)

aj = arcjet_sync(
    key=arcjet_key,
    rules=[
        detect_prompt_injection(mode=Mode.LIVE),
        detect_sensitive_info(
            mode=Mode.LIVE,
            deny=[
                SensitiveInfoEntityType.EMAIL,
                SensitiveInfoEntityType.CREDIT_CARD_NUMBER,
                SensitiveInfoEntityType.PHONE_NUMBER,
            ],
        ),
        detect_bot(mode=Mode.LIVE, allow=["CURL"]),
        token_bucket(
            characteristics=["userId"],
            mode=Mode.LIVE,
            refill_rate=5,
            interval=10,
            capacity=10,
        ),
    ],
)


@app.post("/chat")
def chat():
    body = request.get_json()
    message = body.get("message", "") if body else ""

    decision = aj.protect(
        request,
        requested=5,
        characteristics={"userId": "user_123"},
        detect_prompt_injection_message=message,  # scan for prompt injection
        sensitive_info_value=message,  # scan for PII before sending to LLM
    )

    if decision.is_denied():
        status = 429 if decision.reason_v2.type == "RATE_LIMIT" else 403
        return jsonify(error="Denied"), status

    reply = chain.invoke({"message": message})
    return jsonify(reply=reply)

Arcjet Guard

arcjet.guard is a lower-level API designed for AI agent tool calls and background tasks where there is no HTTP request object. It gives you fine-grained, per-call control over rate limiting, prompt injection detection, content moderation, sensitive information detection, and custom rules.

How it differs from `arcjet` / `arcjet_sync`

	`arcjet` / `arcjet_sync`	`arcjet.guard`
Designed for	HTTP request protection	AI agent tool calls, background jobs
Request object	Required (`protect(request, ...)`)	Not needed
Rule binding	Rules configured once, input via `protect()` kwargs	Rules configured as classes, called with input per invocation
Rate limit key	IP or `characteristics` dict	Explicit `key` string (SHA-256 hashed before sending)
Custom rules	Not supported	`LocalCustomRule` with typed config/input/data

Installation

arcjet.guard is included in the arcjet package — no extra install required.

Quick start

import os
from arcjet.guard import (
    launch_arcjet,  # async — use launch_arcjet_sync for sync frameworks
    TokenBucket,
    DetectPromptInjection,
    ModerateContent,
    LocalDetectSensitiveInfo,
)

arcjet_key = os.getenv("ARCJET_KEY")
if not arcjet_key:
    raise RuntimeError("ARCJET_KEY is required")

# Create a single guard client and reuse it.
# Guard never reads ARCJET_KEY itself — pass key= explicitly (same as JS Guard).
aj = launch_arcjet(key=arcjet_key)

# Configure rules once at startup
user_limit = TokenBucket(
    refill_rate=100,
    interval_seconds=60,
    max_tokens=1000,
)
prompt_scan = DetectPromptInjection()
moderate = ModerateContent()
sensitive = LocalDetectSensitiveInfo(deny=["EMAIL", "CREDIT_CARD_NUMBER"])


# At call time, bind input and guard
async def handle_tool_call(user_id: str, message: str):
    decision = await aj.guard(
        label="tools.weather",
        rules=[
            user_limit(key=user_id, requested=5),
            prompt_scan(message),
            moderate(message),
            sensitive(message),
        ],
    )

    if decision.conclusion == "DENY":
        raise RuntimeError(f"Blocked: {decision.reason}")

    if decision.has_failed_open():
        raise RuntimeError("Guard unavailable; refusing to run the tool")

    # Allowed by a fully evaluated policy; safe to proceed.

Guard surfaces, extras, and propagation

Arcjet Guard protects code that has no HTTP request — a tool call, a worker, an MCP handler. Four surfaces do that, and which one you want depends on what you are holding when the effect happens.

You have	Use	Needs	Blocks a call?
Any Python callable — a worker, an MCP handler, a job	`guard_action` / `guard_action_sync`	`arcjet`	Yes
A LangChain `BaseTool` you call yourself	`guard_tool`	`arcjet[langchain]`	Yes
An agent built with `create_agent`, whose tool calls the model chooses	`ArcjetMiddleware`	`arcjet[langchain-agents]`	Yes
A chain or agent you want to observe	`ArcjetCaptureHandler`	`arcjet[langchain]`	No — records only

Two rules of thumb. If you can name the tool at wiring time, guard_tool is the smaller change — it returns something that is the tool, so nothing downstream changes. If the model picks the tool and you want one policy per tool name, use the middleware. They compose: a guarded tool called from inside a guarded agent evaluates once per policy, and both land on the same Sequence.

ArcjetCaptureHandler never blocks anything. LangChain's callback dispatch ignores what a handler returns, so a callback cannot deny a call — it is the right home for capture and the wrong home for policy.

Installation extras unlock specific surfaces:

arcjet[langchain] — guard_tool and the capture handlers (ArcjetCaptureHandler and ArcjetAsyncCaptureHandler). Depends on langchain-core only — no LangGraph.
arcjet[langchain-agents] — adds ArcjetMiddleware, and pulls in LangChain and LangGraph.

These are separate because LangGraph is large and optional; many applications use tools without agents. Importing arcjet.guard.langchain never loads LangGraph — that happens only when you reference ArcjetMiddleware or ToolPolicy, and without the extra installed you get an error naming it.

Everything comes from one import path:

from arcjet.guard.langchain import (
    ArcjetCaptureHandler,
    ArcjetMiddleware,
    ToolPolicy,
    guard_tool,
)

Fail-closed default — The checkpoint surfaces (guard_action, guard_action_sync, guard_tool, ArcjetMiddleware) default to denying when evaluation is unavailable. This is Arcjet's one documented divergence from the platform-wide fail-open convention. Note the scope: the core guard() call still fails open, returning an ALLOW for which has_failed_open() is True, as does the request protection API (arcjet / arcjet_sync). It is the surfaces that wrap a consequential effect which fail closed:

on_guard_error="deny" (default) — If Guard is unavailable or evaluation fails, the guarded action blocks with ArcjetUnavailableError.
on_guard_error="allow" — Opt out: if Guard is unavailable, the action runs anyway. Use this only where availability matters more than enforcement.

A DENY conclusion always blocks, with ArcjetDeniedError, regardless of on_guard_error. The two errors are deliberately distinct: a denial means policy evaluated and said no, while unavailability means the check never happened. on_guard_error opts out of failing closed on an unevaluated policy — never out of an evaluated denial.

Correlation and Sequences — Trace requests across boundaries using a correlation ID:

from arcjet.guard import arcjet_sequence, current_correlation_id

# Open a sequence with an explicit ID
with arcjet_sequence(correlation_id="request-id-123"):
    # All guard() and capture() calls inside see this ID
    payload = {"correlation_id": current_correlation_id()}  # "request-id-123"
    # Pass payload to a worker...

# Outside the sequence, ID is None
assert current_correlation_id() is None


# Worker resumes the sequence by reopening it. Nothing is inherited across a
# broker or a bare thread, so the ID has to arrive as data.
def worker(payload):
    with arcjet_sequence(correlation_id=payload["correlation_id"]):
        # Every decision and capture in here lands on the originating
        # request's Sequence rather than starting one of its own.
        send_the_thing()

Precedence for ID selection: explicit correlation_id argument → RunnableConfig["configurable"]["arcjet_correlation_id"] (then its ["metadata"]) → ambient context var → none. A checkpoint outside any sequence still evaluates policy; it just produces decisions that join no Sequence. arcjet_sequence() called with no ID generates a sortable one, which is the right answer only for a genuinely new entrypoint — derive it from a session the caller already has when you can, or you build a Sequence nobody goes looking for.

Inside an agent, pass the ID through the config instead. Both the middleware and any guarded tool below it read the same key, so one run produces one Sequence rather than splitting between them:

await agent.ainvoke(
    {"messages": [...]},
    config={"configurable": {"arcjet_correlation_id": session.id}},
)

The config wins over an enclosing arcjet_sequence, and configurable is checked before metadata. LangChain's own run_id is deliberately not used: a Sequence should be joinable from the session a human would go looking for, not from an ID the framework mints per run.

Propagation across boundaries — Context variables do not cross process or executor boundaries. To carry a correlation ID:

Boundary	Propagates?	Solution
`asyncio.create_task()`	Yes	No action needed
`asyncio.to_thread()`	Yes	No action needed
`ThreadPoolExecutor.submit()`	No	Use `copy_context()`
`loop.run_in_executor()`	No	Use `copy_context()`
`ProcessPoolExecutor`	Not reliably	A `fork` start method happens to inherit; `spawn` and `forkserver` do not. Never depend on it — serialize the ID as data
Generators, async generators	Only while consumed inside	A generator does not carry the context it was created in, so one consumed after the block exits sees nothing. Read the ID before yielding
Celery / RQ / Dramatiq	No	Carry the ID in the job payload

For ThreadPoolExecutor and loop.run_in_executor, use the copy_context() idiom:

from concurrent.futures import ThreadPoolExecutor
from contextvars import copy_context

# Inside a sequence context
ctx = copy_context()
with ThreadPoolExecutor(max_workers=1) as executor:
    # The worker runs with the same correlation context
    result = executor.submit(ctx.run, worker_function).result()

Note: A single Context object cannot be entered twice concurrently. Create a fresh copy_context() for each submission.

Redaction is a named future phase — Sensitive information currently denies rather than redacting. No SDK surface mutates content; a value that fails the sensitive-info check is not rewritten and passed through. Redaction would require an explicit carve-out from the checkpoint ADR's no-mutation rule. Do not assume redaction is coming for free.

Remotely configured policy inputs

Map application values explicitly as server-visible or local. Server inputs are evaluated and retained as policy evidence. Local strings remain in SDK memory; only a correlation digest and local-rule attestation are sent.

from arcjet.guard import local_input, server_input

decision = await aj.guard(
    label="email.sent",
    actor=user_id,
    inputs={
        "recipient": server_input.string(to),
        "subject": local_input.string(subject),
        "content": server_input.string(body),
    },
)

print(decision.policy_evaluation, decision.policy_results)

Rules are optional. Passing no rules (or rules=[]) still calls Guard and sends the label, actor, and policy inputs, so a remotely configured policy can protect the action. An empty rules list does not mean “allow without checking.” When a local input triggers an enforced remote-policy rule, the SDK omits raw and server-exposed policy inputs and sends only privacy-safe local evidence to Guard. Guard records and returns the final decision and decision ID.

LangChain tool checkpoints

Install the optional integration with pip install "arcjet[langchain]", then wrap a tool immediately before execution:

from arcjet.guard import local_input, server_input
from arcjet.guard.langchain import guard_tool

guarded_send_email = guard_tool(
    guard=aj,
    tool=send_email_tool,
    action="email.sent",
    on_guard_error="deny",  # default — blocks if Guard is unavailable
    actor=lambda config: config["configurable"]["user_id"],
    inputs=lambda arguments, _config: {
        "recipient": server_input.string(arguments["to"]),
        "subject": local_input.string(arguments["subject"]),
        "content": server_input.string(arguments["body"]),
    },
)

The guarded tool advertises the same schema as the tool it wraps and delegates to it, so the model is told what it would have been told without the guard. Every way of calling the tool is a checkpoint — invoke(), ainvoke(), run(), arun(), streaming and batching, and the tool's own func — and each call is evaluated exactly once. LangChain normalizes a call before the checkpoint reads it, so a policy resolver sees the config the tool runs with, including one a chain passed down without the caller re-threading it.

The guarded tool is an instance of the wrapped tool's class, so application code and LangChain itself keep taking the same branch when they check a tool's concrete type. It is a generated subclass, created once per tool class and kept for the life of the process: a tool class that hooks __init_subclass__ to register or validate its subclasses sees that subclass once, at the first guard_tool() call for it, and a registry keyed by class name gains an entry for it. A trace still shows the tool's own name; it is repr() and the class name that show ArcjetGuarded<ClassName>.

A tool that keeps its own state under one of the names the guard uses for its own — _arcjet_state — is refused by guard_tool(), because a guarded tool is an instance of the tool's class and there is nowhere else for either to live. Rename the tool's attribute.

A blocked call is reported to the tool's callbacks the way LangChain reports the outcome: a denial the tool's handle_tool_error converts is a run that ends with the handled content, and anything else is a start followed by an error. So a blocked call appears on a trace rather than leaving a gap, and a handled denial is not counted as a failure. What the handler returns is formatted by LangChain's own code, so a denial reaches the model shaped exactly as the tool's own error would have been.

If a resolver fails, Guard still sees the call — the decision is made without that input rather than not made at all — and on_guard_error decides whether the call may run.

Configure the tool before you guard it. The guarded tool carries a copy of the tool's state, but the wrapped tool is what executes, so anything you change on the guarded tool afterwards does not reach the call — callbacks, tags, handle_tool_error, args_schema, response_format, and any other field. A blocked call is reported from the wrapped tool too, so allowed and blocked calls agree with each other.

The same applies to a method: if the tool's class has a helper that configures it — bind_user(), say — calling that on the guarded tool sets the value on the guarded tool, and the wrapped tool still runs without it. Call it before guarding, or call it on the tool you still hold.

response_format is worth calling out because ignoring it does more than nothing: a tool that returns (content, artifact) runs with the wrapped tool's format, so the tuple is JSON-encoded into the message content and the artifact is dropped.

args_schema follows that same rule, which matters if you want to hide an argument from the model. Narrow the tool before you guard it, so the wrapped tool parses against the narrow schema and the hidden argument is discarded rather than reaching the tool body. Narrowing the guarded tool instead changes nothing at all, because every schema question is answered by the tool it wraps.

class PublicEmailArgs(BaseModel):
    to: str  # `internal_note` is deliberately absent


send_email_tool.args_schema = PublicEmailArgs  # narrow first, then guard
guarded_send_email = guard_tool(guard=aj, tool=send_email_tool, action="email.sent")

Note that "discarded" is not "rejected": pydantic ignores an unknown field by default, so a caller that sends internal_note anyway gets a successful call with the field dropped, not an error. To reject it, give the narrow schema model_config = ConfigDict(extra="forbid"), or bind a rule to the argument.

Narrow before guarding for the same reason you configure before guarding: changing the wrapped tool's schema afterwards leaves the two disagreeing about what to advertise, for a tool built with Tool(...) and no schema of its own.

Warning

Pickle executes arbitrary code as it loads. Load a pickled tool only from a source you control — your own worker queue or process pool — and never from a user, a network peer, or shared storage anyone else can write to. A guarded tool is not a safe transport format for untrusted input, and the checkpoint does not protect the load: the code runs before any rule does.

A guarded tool can be pickled if the tool it wraps can and its resolvers can, which is what sending tools to a worker process needs. Fields set on the guarded tool survive the round trip. Resolvers are pickled as you gave them, so a lambda or a closure makes the tool unpicklable — use a module-level function if the tool has to cross a process boundary. The client is not pickled with it — it cannot cross a process boundary, and pickling it would write your site key into whatever the pickle is stored in — so call register_arcjet() in the receiving process before loading the tool.

The core guard() API and the LangChain helper have intentionally different defaults when evaluation is unavailable:

API	Default when Guard is unavailable	How to change it
`guard()` (core)	Allow (fail open), with `has_failed_open()` returning `True`	Gate manually on `has_failed_open()`
`guard_tool()`	Block (fail closed)	Set `on_guard_error="allow"`

For guard_tool(), unavailable means either that the pre-execution checkpoint raised while resolving the actor or policy inputs, or calling Guard, or that Guard returned an ALLOW decision whose has_failed_open() is True. The latter can result from a deadline, response parse failure, local rule failure, missing decision, or server-returned rule error—not only an Arcjet Cloud outage.

With the default on_guard_error="deny", the wrapped tool does not execute and ArcjetToolUnavailableError is raised. This is distinct from ArcjetToolDeniedError, which represents a real DENY decision and carries that decision. Handle an unavailable evaluation as an operational failure that may warrant alerting or retrying; do not treat it as a policy denial. Set on_guard_error="allow" only at call sites where availability matters more than enforcement, such as a read-only lookup.

Guarding an agent's tool calls

When the model chooses the tool, guard the agent instead of each tool. Install arcjet[langchain-agents], then give create_agent the middleware:

from langchain.agents import create_agent

from arcjet.guard.langchain import ArcjetMiddleware, ToolPolicy

agent = create_agent(
    model="openai:gpt-4o",
    tools=[send_email, search_orders],
    middleware=[
        ArcjetMiddleware(
            policies={"send_email": ToolPolicy(action="email.sent")},
            tools=[send_email, search_orders],
        )
    ],
)

Tools without a policy pass through unguarded, so you protect the ones that do something consequential and leave the rest alone.

Pass tools= the same sequence you gave create_agent. A policy is matched by tool name, so without it a typo — or a renamed @tool function — leaves that tool unguarded and looks exactly like a healthy allow. Given the tools, a key naming none of them is refused where you wrote it.

The client is optional. Without guard=, the checkpoint uses whatever you registered with register_arcjet(), which is what you want in an application that registers once at startup.

Sync usage

For Flask, Django, or other sync frameworks, use launch_arcjet_sync:

from arcjet.guard import launch_arcjet_sync, TokenBucket

aj = launch_arcjet_sync(key=arcjet_key)
user_limit = TokenBucket(refill_rate=10, interval_seconds=60, max_tokens=100)


def handle_tool_call(user_id: str):
    decision = aj.guard(
        label="tools.weather",
        rules=[user_limit(key=user_id)],
    )

    if decision.conclusion == "DENY":
        raise RuntimeError("Rate limited")

    if decision.has_failed_open():
        raise RuntimeError("Guard unavailable; refusing to run the tool")

Rate limiting

Token bucket, fixed window, and sliding window algorithms are available. Configure the rule once, then call it with a key (and optional requested token count) for each invocation:

Token bucket

from arcjet.guard import TokenBucket

user_limit = TokenBucket(
    refill_rate=100,  # tokens added per interval
    interval_seconds=60,  # seconds between refills
    max_tokens=1000,  # maximum bucket capacity
)

# At call time:
decision = await aj.guard(
    label="tools.weather",
    rules=[user_limit(key=user_id, requested=5)],
)

Fixed window

from arcjet.guard import FixedWindow

team_limit = FixedWindow(
    max_requests=1000,
    window_seconds=3600,
)

decision = await aj.guard(
    label="api.search",
    rules=[team_limit(key=team_id)],
)

Sliding window

from arcjet.guard import SlidingWindow

api_limit = SlidingWindow(
    max_requests=500,
    interval_seconds=60,
)

decision = await aj.guard(
    label="api.query",
    rules=[api_limit(key=user_id)],
)

Prompt injection detection

from arcjet.guard import DetectPromptInjection

prompt_scan = DetectPromptInjection()

decision = await aj.guard(
    label="tools.weather",
    rules=[prompt_scan(user_message)],
)

if decision.conclusion == "DENY":
    print("Prompt injection detected")

result = prompt_scan.result(decision)
if result and result.billing:
    print(result.billing.unit, result.billing.count)

Guard billing is optional. Prompt injection usage is reported in tokens, while content moderation usage is reported in text_units (text chunks), so always inspect billing.unit rather than assuming the unit.

Content moderation

from arcjet.guard import ModerateContent

moderate = ModerateContent()

decision = await aj.guard(
    label="tools.chat",
    rules=[moderate(user_message)],
)

if decision.conclusion == "DENY":
    print("Harmful content detected")

result = moderate.result(decision)
if result:
    print(result.detected)
    if result.billing:
        print(result.billing.unit, result.billing.count)

experimental_ModerateContent remains as a deprecated alias for ModerateContent.

The result reports detected and optional billing only — not per-category scores.

Sensitive information detection

Detects PII locally — the raw text never leaves the SDK. The default backend detects EMAIL, PHONE_NUMBER, IP_ADDRESS, CREDIT_CARD_NUMBER.

from arcjet.guard import LocalDetectSensitiveInfo

sensitive = LocalDetectSensitiveInfo(
    deny=["EMAIL", "CREDIT_CARD_NUMBER"],
)

decision = await aj.guard(
    label="tools.send_email",
    rules=[sensitive(user_input)],
)

For additional entity types (names, addresses, SSN, etc.), install arcjet[sensitive-info-rampart] and pass the on-device Rampart backend:

from arcjet.guard import LocalDetectSensitiveInfo
from arcjet_sensitive_info_rampart import rampart

sensitive = LocalDetectSensitiveInfo(deny=["GIVEN_NAME", "SSN"], backend=rampart())

Listing a backend-only type without a supporting backend raises.

Custom rules

Define typed custom rules that run locally. Subclass LocalCustomRule and override evaluate (sync) or evaluate_async (async):

from typing import TypedDict
from arcjet.guard import LocalCustomRule, CustomEvaluateResult


class TopicConfig(TypedDict):
    blocked_topic: str


class TopicInput(TypedDict):
    topic: str


class TopicData(TypedDict):
    matched: str


class TopicBlockRule(LocalCustomRule[TopicConfig, TopicInput, TopicData]):
    def evaluate(
        self,
        config: TopicConfig,
        input: TopicInput,
    ) -> CustomEvaluateResult:
        if input["topic"] == config["blocked_topic"]:
            return CustomEvaluateResult(
                conclusion="DENY",
                data={"matched": input["topic"]},
            )
        return CustomEvaluateResult(conclusion="ALLOW")


rule = TopicBlockRule(config={"blocked_topic": "weapons"})
inp = rule(data={"topic": user_topic})
decision = await aj.guard(rules=[inp], label="content")

# Access typed result
r = inp.result(decision)
if r and r.conclusion == "DENY":
    print(f"Blocked topic: {r.data['matched']}")

Per-rule results

Both the configured rule and the bound input provide typed result accessors:

user_limit = TokenBucket(refill_rate=10, interval_seconds=60, max_tokens=100)
inp = user_limit(key=user_id, requested=5)

decision = await aj.guard(label="tools.weather", rules=[inp])

# From the bound input (matches exact invocation)
r = inp.result(decision)
if r:
    print(r.remaining_tokens, r.max_tokens)

# From the configured rule (matches all invocations of this rule)
r = user_limit.result(decision)

# Check only denied results
denied = inp.denied_result(decision)
if denied:
    print(f"Rate limited — resets at {denied.reset_at_unix_seconds}")

Decision API

decision = await aj.guard(label="tools.weather", rules=[...])

# Layer 1: conclusion and reason
decision.conclusion  # "ALLOW" or "DENY"
decision.reason  # "RATE_LIMIT", "PROMPT_INJECTION", "MODERATE_CONTENT", "SENSITIVE_INFO", "CUSTOM", "ERROR", etc.

# Layer 2: error/warning detection
decision.has_failed_open()  # True if ALLOW only because a rule/decision could not be processed (fail-closed gate)
decision.error_results()  # Results that errored (rules or the decision that could not be processed)
decision.warnings  # Decision-level diagnostics (e.g. an invalid metadata key that was stripped)

# Layer 3: per-rule results (see "Per-rule results" above)
for result in decision.results:
    print(result.type, result.conclusion)

`guard()` parameter reference

Parameter	Type	Description
`rules`	`Sequence[RuleWithInput]`	Bound SDK rule inputs (optional; defaults to empty for policy-only calls)
`label`	`str`	Label identifying this guard call (required)
`actor`	`str \| None`	Actor used by remote policy selection/evaluation
`inputs`	`PolicyInputMap \| None`	Typed server-visible or local remote-policy inputs
`metadata`	`Metadata \| None`	Structured metadata — see Metadata
`correlation_id`	`str \| None`	Opaque id correlating this call with other `guard()`/`protect()` calls

Metadata

guard(), protect(), and every guard rule accept metadata: a mapping of string keys to any JSON-serializable value, including nested objects and arrays. It is attached to the decision for correlation and analytics.

decision = await aj.guard(
    label="tools.weather",
    rules=[user_limit(key=user_id)],
    metadata={
        "user": {"id": user_id, "plan": "pro"},
        "tool_name": "get_weather",
        "duration_ms": 160,
        "success": True,
    },
)

Each top-level value is JSON-encoded by the SDK and stored verbatim, so exact integers and value formatting survive. Server-enforced limits:

Limit	Value	Over the limit
Top-level keys	128	Extra keys dropped
Serialized bytes / value	4 KiB	That key dropped
Nesting depth / value	10	That key dropped
Key names	letters, digits, `-`, `.`, `_`	That key dropped

Nothing here can fail a call or change a decision — metadata is excluded from fingerprinting and from the decision cache key. Every dropped key is reported: server-side drops arrive on decision.warnings, one per key. Keys the SDK itself could not encode (a datetime, a set, NaN, a circular reference) are collected into a single warning naming them all, added to decision.warnings and reported to the server. For protect(), which has no warnings channel on its Decision, that warning is logged at WARNING instead.

Metadata is untrusted and is not redacted — do not put secrets or PII in it.

Some limits are the SDK's own, not the server's. The SDK drops keys once one request's metadata exceeds 768 KiB in total (keys plus JSON-encoded values, counted before compression). That ceiling sits well above anything the server would accept — its own caps allow roughly 512 KiB in a single map — and exists only so oversized metadata cannot push a request past the 1 MiB protocol limit, where it would be rejected outright and fail open.

Two behaviours differ between the Python and JavaScript SDKs:

Integer precision. Python integers are arbitrary-precision and are sent verbatim, so a value past 2^53 survives exactly. The JavaScript SDK cannot do this — its numbers are IEEE-754 doubles before they reach the wire — so send such values as strings if both SDKs must agree.
Objects with a toJSON() method, including JavaScript Date, are serialized by that method in the JS SDK. Python has no equivalent protocol, so a datetime (or any other non-JSON type) is dropped with a warning. Convert explicitly — datetime.isoformat() — if both SDKs must agree.

Rule-level metadata is merged with guard()-level metadata shallowly: a duplicate key's whole value is replaced, never deep-merged.

DRY_RUN mode

All guard rules accept a mode parameter. Use "DRY_RUN" to evaluate rules without blocking:

user_limit = TokenBucket(
    refill_rate=10,
    interval_seconds=60,
    max_tokens=100,
    mode="DRY_RUN",
)

Recording what happened with `capture()`

guard() decides whether something is allowed. capture() records that it happened. Use it for the actions you want to see in a security trace but do not want to gate — a refund issued, a document exported, a tool call completed.

decision = await aj.guard(label="refund", rules=[inp])
if decision.conclusion == "ALLOW":
    refund_id = issue_refund(...)

    aj.capture(
        action="refund.issued",
        correlation_id=workflow_id,  # ties this to other calls in the workflow
        decision_id=decision.id,  # ties it to the decision above
        metadata={"amount_cents": 4999, "invoice": {"id": "inv_123"}},
    )

capture() returns immediately and is not awaited, even on the async client. Events are queued and sent in the background, batched together.

It is best-effort and never affects a decision:

It never raises. A bad field is dropped and the rest of the event is sent; an event with no usable action is dropped entirely.
Under sustained load or a failing backend, events are dropped rather than slowing your request down. A failed send is never retried.
Nothing is dropped silently. Drops are reported through the arcjet logger with a stable code — AJ3001 (queue full), AJ3002 (send failed), AJ3003 (flush deadline). The arcjet logger is already at WARNING, so you only need to attach a handler to see them.

Repeats of the same code are coalesced for a minute and the suppressed count is reported with the next line for that code, or by the next flush(). A burst that ends without either will under-report its total — the figure is a count of events seen, not a guaranteed total.

Pass your own logger to receive every diagnostic uncoalesced, which is what you need to keep a metric of dropped events:
```
aj = launch_arcjet_sync(key=arcjet_key, logger=my_logger)
```
Each record carries code and count attributes alongside the message, so a handler can route or count on them without parsing text.

Do not put secrets or PII in metadata; it is stored as untrusted data.

Delivering events before shutdown

Delivery is asynchronous, so events queued as your process exits may never be sent — the sync worker is a daemon thread and will not hold the interpreter open. Call flush() at a shutdown point:

# Async (FastAPI lifespan, or any async teardown)
await aj.flush()

# Sync (Flask teardown, atexit, or the end of a script)
arcjet_sync_guard.flush()

flush() waits up to timeout_ms (default 1000) for the events outstanding when you called it. On expiry, queued events are dropped and a request already on the wire is abandoned — not cancelled, so it may still arrive, and nothing will tell you either way. Both are counted in the AJ3003 report.

Events captured while a flush is waiting are not its responsibility and survive its deadline, so calling flush() per request in a concurrent server cannot discard another request's telemetry.

There is no close(): a client holds no connection of its own to release, so flushing is the only shutdown step that changes what gets delivered.

Registering a client (optional)

Passing the client explicitly is the recommended path, and everything above does exactly that. Registration is a shortcut for the case it cannot cover: code too deep in an application to be handed a client, where capture() is often most useful.

launch_arcjet() never touches global state. Registering is always a separate, explicit call:

# wherever your application starts up
import os

from arcjet.guard import launch_arcjet, register_arcjet

register_arcjet(launch_arcjet(key=os.environ["ARCJET_KEY"]))

capture() is then importable on its own and reaches the registered client:

# deep in application code — nothing was passed down here
from arcjet.guard import capture


async def refund(invoice_id: str) -> None:
    await issue_refund(invoice_id)
    capture(action="refund.issued", metadata={"invoice": invoice_id})

Sync and async

capture() is one function for both client flavors, because it queues and returns on each of them. guard() and flush() cannot be, so they come in pairs matching launch_arcjet / launch_arcjet_sync:

Registered client	Guard	Flush
`launch_arcjet()`	`await guard(...)`	`await flush()`
`launch_arcjet_sync()`	`guard_sync(...)`	`flush_sync()`

Registration accepts either and does not record which, so calling the wrong one is not something a type checker can catch. It fails open and reports AJ3007 on the registered client's logger.

What happens with nothing registered

guard() and guard_sync() return a fail-open ALLOW carrying an error result, so a caller that inspects the decision can see that no policy ran. They do not raise.

decision = await guard([limit(key=user_id)], label="refund")

if decision.has_failed_open():
    # No rule was evaluated. Treat this as "policy did not run", not as a pass.
    ...

capture() drops the event silently, and the flush() variants return. Nothing is logged: the client that would have carried a logger is the thing that is missing, so the only available sink would be an unconfigurable warning on a request path.

Registering twice, and unregistering

Registration is guarded. A second client does not displace the first — the attempt is reported as AJ3004 on the incumbent's logger, so a library or a stray second launch_arcjet() cannot quietly redirect an application's telemetry to a different key. Registering the client that is already registered is a silent no-op.

unregister_arcjet() takes no argument and clears whatever is there. The cost is that anything calling it clears the application's client and every free call afterwards fails open, so libraries should not call it — they take a client explicitly. That is a convention, not something the SDK enforces.

The registration is a module-level global, so it is visible from every thread and every event loop. It is deliberately not a contextvars.ContextVar: a context variable set at startup is invisible inside worker threads, which is exactly how Flask, Django and other WSGI servers run request handlers, so registration would appear to work in development and silently do nothing in production.

Testing

arcjet.guard.testing registers an in-memory client that records calls and talks to nothing:

from arcjet.guard import capture
from arcjet.guard.testing import register_test_client


async def test_refund_captures_an_event():
    with register_test_client() as arcjet:
        await refund("inv_1")

        assert arcjet.captures[0].action == "refund.issued"

It drives every guard surface — the free calls, guard_tool, the middleware and the capture handlers alike — because each identifies a client by the shape of its guard() rather than by its class:

from arcjet.guard.langchain import guard_tool
from arcjet.guard.testing import register_test_client


def test_sending_an_email_is_guarded():
    with register_test_client() as arcjet:
        guarded = guard_tool(
            guard=arcjet,
            tool=send_email,
            action="email.sent",
            on_guard_error="allow",
        )
        guarded.invoke({"to": "a@b.c", "subject": "hi"})

        assert [call.label for call in arcjet.guards] == ["email.sent"]

Pass on_guard_error="allow" unless the test is asserting a denial: the recorder answers a fail-open decision, which the default treats as an unevaluated policy and refuses.

The with block unregisters the client on the way out, including when the test fails part-way through. Note the await: the capture happens wherever the code under test reaches it, so a test that forgets to await an async function asserts before the event exists.

Usually this belongs in a fixture:

import pytest
from arcjet.guard.testing import register_test_client


@pytest.fixture
def arcjet():
    with register_test_client() as client:
        yield client

register_test_client() raises if a client is already registered, which surfaces a leak from an earlier test rather than letting this one assert against the wrong recorder. unregister() is also available for teardown that cannot use with, and only clears the registration if it is still this client.

Each recorded capture goes through the same validation and metadata encoding as a real capture(), so a call the real client would drop is not recorded here either, and anything the SDK rewrote is on capture.warnings. Recording itself is synchronous — once the code under test reaches capture(), the event is there with no flushing or waiting.

The test client answers both guard() and guard_sync(), so it does not care which flavor your application uses. It records the call and returns a fail-open ALLOW, because no rule actually ran. It is not a mock server and does not let you stub per-rule verdicts.

guard_tool() accepts it too — it identifies a client by the shape of its guard(), not by its class. Because the recorder answers a fail-open decision, pass on_guard_error="allow" unless the test is asserting the denial.

Best practices

Single-instance pattern

Create one Arcjet client at startup and reuse it across all requests:

# Good — one instance, created once at startup
aj = arcjet(key=arcjet_key, rules=[...])


# Bad — new instance per request wastes resources
@app.get("/")
async def index(request: Request):
    aj = arcjet(key=arcjet_key, rules=[...])  # don't do this

DRY_RUN mode for testing

Use Mode.DRY_RUN to test rules without blocking traffic. Decisions are logged but requests are allowed through:

aj = arcjet(
    key=arcjet_key,
    rules=[
        detect_bot(mode=Mode.DRY_RUN, allow=[]),
        token_bucket(mode=Mode.DRY_RUN, refill_rate=5, interval=10, capacity=10),
    ],
)

Proxy configuration

When running behind a load balancer or reverse proxy, configure trusted IPs so Arcjet resolves the real client IP from X-Forwarded-For:

aj = arcjet(
    key=arcjet_key,
    rules=[...],
    proxies=["10.0.0.0/8", "192.168.0.1"],
)

Outbound HTTP proxy

If your environment requires outbound requests to the Arcjet API to go through a forward proxy (e.g. Squid), set the standard proxy environment variables. The SDK honors them automatically — no code changes required:

export HTTPS_PROXY="http://proxy.example.com:3128"
# Optional: comma-separated hosts that should bypass the proxy
export NO_PROXY="decide.arcjet.com,localhost"

HTTP_PROXY, HTTPS_PROXY, and NO_PROXY are all supported (the lower-case variants work too). Because the Arcjet API is reached over HTTPS, HTTPS_PROXY is the relevant variable for proxying Arcjet traffic. NO_PROXY accepts a comma-separated list of hostnames to bypass, or * to disable proxying entirely. Since Arcjet is reached by hostname, list the hostname (e.g. decide.arcjet.com) to bypass it.

Async vs. sync client

Use arcjet (async) with FastAPI and other async frameworks. Use arcjet_sync with Flask and other sync frameworks:

from arcjet import arcjet, arcjet_sync

# Async — for FastAPI, Starlette, etc.
aj_async = arcjet(key=arcjet_key, rules=[...])
decision = await aj_async.protect(request)

# Sync — for Flask, Django, etc.
aj_sync = arcjet_sync(key=arcjet_key, rules=[...])
decision = aj_sync.protect(request)

`protect()` parameter reference

All parameters are optional keyword arguments passed alongside the request:

Parameter	Type	Used by
`requested`	`int`	Token bucket rate limit
`characteristics`	`dict[str, Any]`	Rate limiting (pass values for keys declared in rule config)
`detect_prompt_injection_message`	`str`	Prompt injection detection
`sensitive_info_value`	`str`	Sensitive info detection
`email`	`str`	Email validation
`filter_local`	`dict[str, str]`	Request filters (`local.*` fields)
`metadata`	`Metadata`	Structured metadata — see Metadata
`correlation_id`	`str`	Opaque id correlating this call with other `protect()`/`guard()` calls
`ip_src`	`str`	Manual IP override (advanced)

Decision response

decision = await aj.protect(request)

# Top-level checks
decision.is_denied()  # True if any rule denied the request
decision.is_allowed()  # True if all rules allowed the request
decision.is_error()  # True if Arcjet encountered an error (fails open)

# reason_v2.type values: "BOT", "RATE_LIMIT", "SHIELD", "EMAIL", "ERROR", "FILTER"
if decision.reason_v2.type == "RATE_LIMIT":
    print(decision.reason_v2.remaining)  # tokens/requests remaining
elif decision.reason_v2.type == "BOT":
    print(decision.reason_v2.denied)  # list of denied bot names
    print(decision.reason_v2.spoofed)  # list of spoofed bot names

# Per-rule results (for granular handling)
for result in decision.results:
    print(result.reason_v2.type, result.is_denied())

Error handling

Arcjet is designed to fail open — if the service is unavailable, requests are allowed through. Check for errors explicitly if your use case requires it:

decision = await aj.protect(request)

if decision.is_error():
    # Arcjet service error — fail open or apply fallback policy
    pass
elif decision.is_denied():
    return JSONResponse({"error": "Denied"}, status_code=403)

Support

This repository follows the Arcjet Support Policy.

Security

This repository follows the Arcjet Security Policy.

Compatibility

Packages maintained in this repository support CPython 3.10+ on macOS, Windows, and glibc Linux (Debian, Ubuntu, RHEL, and *-slim / manylinux container images).

Two runtime dependencies ship native code:

wasmtime — local WASM rule evaluation
pyqwest — HTTP client used by connect-python

Those packages publish musllinux wheels today, so pip install arcjet often succeeds on Alpine without a compiler. Alpine/musl is still not a supported target. A future wheel gap forces a source build (Rust plus a C toolchain), which is what broke python3.10-alpine images after local analysis landed. Prefer a glibc image such as python:3.10-slim or astral/uv:python3.10-trixie-slim.

License

Licensed under the Apache License, Version 2.0.

Name		Name	Last commit message	Last commit date
.devcontainer		.devcontainer
.github/workflows		.github/workflows
docs		docs
examples		examples
sensitive-info-rampart		sensitive-info-rampart
src/arcjet		src/arcjet
tests		tests
tools		tools
.gitignore		.gitignore
.python-version		.python-version
AGENTS.md		AGENTS.md
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE		LICENSE
README.md		README.md
justfile		justfile
pyproject.toml		pyproject.toml
renovate.json		renovate.json
uv.lock		uv.lock
witgen.toml		witgen.toml

Uh oh!

Latest commit

History

Folders and files

Repository files navigation

arcjet

Why Arcjet?

Getting started

Install the Arcjet CLI

Quick setup with an AI agent

Manual setup

Get help

Quick start

Features

Which features do I need?

Installation

Prompt injection detection

FastAPI

Flask

Bot protection

FastAPI

Flask

Bot categories

Verified vs. spoofed bots

Rate limiting

Token bucket (recommended for AI)

Fixed window

Sliding window

Sensitive information detection

On-device Rampart backend (more entity types)

Shield WAF

Email validation

Request filters

Block by country

Block VPN and proxy traffic

Custom local fields

IP analysis

LangChain example

FastAPI + LangChain

Flask + LangChain

Arcjet Guard

How it differs from arcjet / arcjet_sync

Installation

Quick start

Guard surfaces, extras, and propagation

Remotely configured policy inputs

LangChain tool checkpoints

Guarding an agent's tool calls

Sync usage

Rate limiting

Token bucket

Fixed window

Sliding window

Prompt injection detection

Content moderation

Sensitive information detection

Custom rules

Per-rule results

Decision API

guard() parameter reference

Metadata

DRY_RUN mode

Recording what happened with capture()

Delivering events before shutdown

Registering a client (optional)

Sync and async

What happens with nothing registered

Registering twice, and unregistering

Testing

Best practices

Single-instance pattern

DRY_RUN mode for testing

Proxy configuration

Outbound HTTP proxy

Async vs. sync client

protect() parameter reference

Decision response

Error handling

Support

Security

How it differs from `arcjet` / `arcjet_sync`

`guard()` parameter reference

Recording what happened with `capture()`

`protect()` parameter reference