README.md

Execution safety for AI agents.
The model guesses. CTRLRun does not.

A Python library that sits between the decision to act and the call that acts.
A consequential action happens at most once, exactly as approved, and leaves a receipt.
When the outcome is unknown, CTRLRun says so instead of guessing.

Runs in production on a single file, or on Postgres across hosts. Apache-2.0.

pip install ctrlrun && ctrlrun demo

What it does

A ticket asks for a €500 refund. The agent calls the refund tool with €5,000, one extra zero. The tool is in its list, the arguments are well formed, and the model is completely confident. Nothing above the call disagrees, because nothing above the call is a check: a tool being callable is not permission to call it with those arguments.

CTRLRun is that check. It reads the arguments about to leave your process and answers what may happen to them. Four rules do the work, and each one is a test in this repository before it is a sentence here.

Exact means exact	Changed arguments need a new approval.
Once stays once	Same effect key, shared store, no repeat.
Unknown means wait	Confirm the outcome before retrying.
Every answer is kept	Requests, decisions and results, refusals included.

The third one is the half people forget. A correct €500 refund commits at the provider and the reply is lost coming back, so the agent retries. Retry libraries, agent frameworks and tool loops collapse this failed into I do not know what happened. CTRLRun keeps them apart: a lost reply is AMBIGUOUS, never FAILED, and a retry against an AMBIGUOUS effect is refused until a human, or a reconcile hook, says what happened.

What ctrlrun demo shows: five failures and five refusals, byte for byte

$ ctrlrun demo
CTRLRun demo — five ways an agent action goes wrong, and what stops it.
Policy: refunds up to €1,000 are autonomous, up to €10,000 need a human, above that are denied.

1. Duplicate effect after a lost response

   refund €500  →  remote commits  →  response lost  →  effect: AMBIGUOUS
   agent retries the same refund
   ✗ BLOCKED — effect may already have committed; blind retry refused
   remote refund calls: 1
   only a human moves it on:  ctrlrun resolve refund:txn_1 --committed|--failed

2. Approval mutation

   agent proposes refund €2,000  →  human approves apr_0aa78e0380ba55d77a601dc782f57095 (bound to the action hash)
   agent executes refund €5,000  →
   ✗ BLOCKED — approved action ≠ requested action (mismatch)

3. Concurrent agents, same effect

   Agent A  reserve refund:txn_123  →  ACQUIRED  →  executes
   Agent B  reserve refund:txn_123  →
   ✗ BLOCKED — already reserved (in_progress)

4. Approval replay

   approval apr_dbc8bc6f06690cdf2e2c55a4e591ef3b used once  →  consumed
   same approval presented again                            →
   ✗ BLOCKED — single-use approval already consumed

5. Authority escalation

   human €100,000 delegable  →  finance agent €25,000  →  support agent €2,000
   support agent's grant: dlg_5f8d41938a3f29972d5489d676cd9edb
   support agent requests €50,000  →
   ✗ BLOCKED — outside the delegated grant (authority_constraint)
   remote refund calls: 0
   finance agent tries to delegate €50,000 under its own €25,000  →  refused (containment: constraints)
   support agent requests €1,500  →  authority permits it, and the policy asks a human (apr_f86eca24dd80206ab5189ccb1b62aa55)
   two axes, and an action needs both: the stricter of the pair wins

Receipts (8): .ctrlrun/demo/receipts.jsonl
Events:       .ctrlrun/demo/events.jsonl

Read them:    CTRLRUN_STATE=.ctrlrun/demo/state.db ctrlrun receipts

Approval and delegation ids are generated per run; everything else is exactly what the demo prints, and a test fails if the two drift apart. No network, no external service, under a second. Run it in your browser with nothing installed.

Where it stops. It does not detect prompt injection: it contains the consequence rather than reading the cause. It cannot promise exactly-once against a remote it does not control, it refuses to knowingly act twice, and it rolls nothing back. Receipts are chained, so an alteration is detected. They are not signed: alteration is not authorship. The badge above means the declared guarantees pass in the setup they ran against, and it does not mean secure, safe, compliant, certified or audited: what the badge means · OWASP-AGENTIC-TOP10.md names the four entries this does not address.

If an agent only reads and answers, you do not need CTRLRun. The moment it can send, pay, refund, delete, deploy, grant, revoke, approve, submit, purchase or cancel, you do.

Use it in three steps

1. Install it.

pip install ctrlrun

2. Write down what the agent may do. One file, ctrlrun.yaml. Cheap to undo is autonomous, handing out power needs a human, destroying the evidence is not an agent action at any size. Anything not listed is denied; there is no default-allow.

schema: ctrlrun.policy/v2

actions:
  # Cheap to undo: the agent does it alone.
  crm.update_record:
    effect: "crm:{record_id}:{field}"
    decision: allow

  # Hands out power: read-only is autonomous, anything above it is a human's call.
  iam.grant_role:
    effect: "grant:{user_id}:{role}"
    rules:
      - when: { role_in: [reader, viewer] }
        decision: allow
      - decision: approve

  # Destroys the evidence: denied, whoever asks.
  audit.log.delete:
    decision: deny

3. Wrap the call that has the consequence, and name the effect it has in the world.

import sys

import ctrlrun


class FakeDirectory:
    calls: list[tuple[str, str]] = []

    def grant(self, user_id: str, role: str) -> dict:
        self.calls.append((user_id, role))
        return {"user": user_id, "role": role, "status": "granted"}


directory = FakeDirectory()


@ctrlrun.protect("iam.grant_role", effect="grant:{user_id}:{role}")
def grant_role(user_id: str, role: str) -> dict:
    return directory.grant(user_id, role)


if __name__ == "__main__":
    with ctrlrun.context(agent="onboarding-agent"):
        print("reader:", grant_role(user_id="u_412", role="reader")["status"])
        try:
            grant_role(user_id="u_412", role="admin")
        except ctrlrun.ApprovalRequired as pending:
            print("admin: a human decides:", pending.request_id)
            with open("request_id.txt", "w") as handle:
                handle.write(pending.request_id)
        else:
            sys.exit("the admin grant ran without a human; the policy is not in force")

The reader grant runs on its own. The admin one stops and names the request a human answers:

reader: granted
admin: a human decides: apr_649156806800a3545de597c028c9dae5

A human answers from the shell, and the grant names the exact action hash it authorizes:

ctrlrun approve "$(cat request_id.txt)"

Present it with ctrlrun.with_approval(request_id) and the agent gets that action and nothing next to it: the same approval spent on owner instead of admin raises ApprovalMismatch. Every attempt, refusals included, leaves a receipt you can read with ctrlrun receipts.

That is the whole integration: a policy file, a decorator, a context, and with_approval to present a grant. Everything else is in the documentation. Protect your first action walks the same path with every output explained · Policy YAML reference · Cookbook: refunds, deploys, IAM, deletions, email, MCP.

How it works

Every protected call, whichever way it arrives, goes through the same six steps.

  normalize  →  decide  →  approve  →  reserve  →  execute  →  record

1. Normalize. The call becomes an Action: a name, canonical arguments (sorted keys, no floats), a resource, the principal. Its SHA-256 is the action hash.
2. Decide. Authority first (may this principal propose this at all?), then policy (how much autonomy does this action get?). Unknown action, missing policy or missing principal is deny.
3. Approve. A human answers against the action hash. The approval is single-use, expires, and matches nothing but that exact action.
4. Reserve. The effect key, refund:txn_1 or namespace:prod-eu:checkout, is taken in one atomic write. A second caller, in another process or on another host, is refused.
5. Execute. Your function runs. Only NotExecuted, raised by you, means FAILED; every other exception and every timeout means AMBIGUOUS.
6. Record. A portable JSON receipt: who, what, decision, approval, effect key, outcome, and the hash of the policy that decided it, chained to the receipt before it.

State lives in SQLite by default, a file with no server and no ops, and the reservation holds across processes rather than merely across threads. Point it at Postgres when more than one host writes: pip install "ctrlrun[postgres]", one URL, the same guarantees graded by the same suite. Prove it in your own setup with ctrlrun verify, which runs the kernel's own failure scenarios against your policy in a scratch store, with no network.

Guarantee	@protect	Gateway	Adapter
Approval binding — An approval is bound to the exact action; a mutated or replayed one is refused.	yes	yes	prevention or attribution, per adapter
One effect, once — One logical effect happens at most once, across threads, processes and hosts.	yes	yes	yes
Unknown is not failed — An unknown outcome is AMBIGUOUS, never FAILED, and blocks a blind retry.	yes	yes	yes
Fail closed — An unknown action, a missing policy or a missing principal is denied.	yes	yes	yes
Authority and delegation — With authority on, every principal needs a grant, and delegation cannot widen one.	yes	yes	yes
Receipts — Every executed action leaves a portable JSON receipt of who, what and outcome.	yes	yes	yes

Three ways to use it

You probably do not need an adapter. @protect covers anything running in this process: a raw model call, a LangChain tool, a hand-rolled loop, a cron job. The gateway covers anything that reaches its tools over MCP, in any language.

You have	Use	Needs
Python in this process	the @protect decorator, shown above	nothing beyond pip install ctrlrun
Tools behind an MCP server, in any language	the gateway: pip install "ctrlrun[gateway]"	one command, no change to agent or server code
A framework with its own approval interrupt	an adapter	the framework to have a human-in-the-loop primitive

An adapter exists for one reason: to route an approve decision through the framework's own interrupt, so a human answers where they already answer. There is never a second place to say yes. ctrlrun-langgraph gives prevention, because the resumption carries the arguments and core re-checks them against the hash. ctrlrun-openai-agents gives attribution, because that SDK records that a call was approved and not what its arguments were. None of the three is only for agents: a worker, a webhook handler and a scheduled job cannot tell a first attempt from a retry either.

The same shape in nine domains

Nothing in CTRLRun knows what a refund is. An action is a name, canonical arguments, an effect key and a resource, and the three questions asked of it are the same whichever domain it came from: how much autonomy does this action get, did a human approve this exact action, and has this effect already happened. Two things carry your domain, and you write both.

• The effect key is the only domain knowledge in the system. It is the string that says two calls are the same real-world consequence: refund:{payment_id}, namespace:{cluster}:{name}, grant:{user_id}:{role}, prescription:{patient_id}:{drug}. Name it well and a retry cannot act twice; leave it out and there is nothing for at most once to be about.
• Conditions are arguments, not amounts. The language is <argument>_<op>, so the same operators read replicas_lte: 10, role_in: [reader, viewer] and to_domain_eq: acme.com as easily as amount_lte. A band is available to a domain that has never issued an invoice.

Domain	Autonomous	A human decides	Never
DevOps	k8s.scale_deployment to 10 replicas	terraform.apply	k8s.delete_namespace
Security operations	firewall.add_deny_rule	firewall.add_allow_rule	edr.disable_protection
Healthcare	appointment.reschedule	patient.export_record	prescription.change_dose
Legal	document.draft_internal	document.file_with_court	contract.execute
HR	pto.approve within a band	payroll.run	employee.delete_record
Insurance	claim.request_documents	claim.approve_payout above a band	policyholder.delete
E-commerce	inventory.adjust within a band	price.update	customer.delete
Public services	eligibility.precheck	benefit.terminate	record.delete
Payments	stripe.refund under €500	stripe.refund above it	stripe.delete_customer

Read any row left to right and it is one rule wearing different nouns. The security row is the one to read twice: adding a deny rule to a firewall is autonomous and adding an allow rule is not, which no amount threshold would have told you. The policy is where your judgement about your domain gets written down; CTRLRun is what makes it hold.

Documentation

ctrlrun.dev is the documentation: concepts, guides, a cookbook, the full reference, and a browser demo that runs with no install.

Start here	Why · Protect your first action · Try it in your browser
The ideas, and doing something with them	Concepts · Guides · Cookbook
MCP	Overview · The gateway in five minutes
Running it for real	Production · Postgres · Recovery · Operations
Every key, flag and error	Reference · FAQ
What holds, and what does not	Threat model · What verify proves · CLAIMS.md, every sentence mapped to its test · How this is built

Contributing

Issues and pull requests are welcome: CONTRIBUTING.md and CODE_OF_CONDUCT.md have the working agreement, and SECURITY.md is how to report a vulnerability. Every claim in this file has a test behind it, so a change to the prose usually means a change to the suite. CHANGELOG.md and https://ctrlrun.dev/docs/ROADMAP say where it is going. Releases carry PyPI provenance attestations from GitHub Actions.

License

Apache-2.0. The enforcement kernel is and will remain fully open source.