Smoke Monkey Desktop

Agentic AI harness — chat with super memory and live visual widgets

A production-grade agent platform: an autonomous coding agent (29 tools, 8 model providers, permission gates, checkpoint/resume) plus conversation memory and document RAG with streamed, cited answers. Comparable to OpenAI Codex, Google Antigravity, and Claude Code's cloud harness. See Agent API Harness.

---

---

Features

Capability	What it does
Agent harness API	Autonomous coding agent over REST + SSE / WebSocket: 29 tools, permission gates, checkpoint/resume, sub-agents, 8 providers (OpenAI, OpenRouter, NVIDIA NIM, Gemini, xAI, OpenCode Zen, OmniRoute, Ollama).
Super memory	Every exchange is embedded into a vector store; durable user facts (preferences, projects) become a long-term profile.
Dynamic visual	The LLM can render animated canvas widgets in the answer stream — flowcharts, sorting demos, algorithm walkthroughs.
Document RAG	Upload PDFs, ask questions, get streamed answers with page-level citations.
Hybrid retrieval	Dense (pgvector) + sparse (Postgres FTS) fused with RRF, cross-encoder reranking, HyDE, multi-query expansion.
Agentic query router	Per-message routing between knowledge base, memory, live web, or direct chat.
User-owned keys	Plug in your own provider keys — encrypted at rest (AES-256-GCM), cached in Redis.

Architecture

flowchart LR
    subgraph clients
        web[Web / Next.js]
        desktop[Tauri Desktop]
    end
    subgraph control
        gateway[API Gateway / NestJS]
        rag[RAG Service / FastAPI]
    end
    subgraph data
        pg[(PostgreSQL + pgvector)]
        redis[(Redis)]
        minio[(MinIO / S3)]
    end
    subgraph ai
        ollama[Ollama local]
        openrouter[OpenRouter / NVIDIA NIM]
    end

    web -- HTTP/SSE --> gateway
    desktop -- HTTP/SSE --> gateway
    gateway -- query + SSE --> rag
    rag --> pg
    rag --> redis
    gateway --> minio
    rag --> ollama
    rag --> openrouter

Loading

Install on macOS

Build from source and install to /Applications in one command (needs node, pnpm, and Xcode Command Line Tools):

./install.sh

• installs workspace deps and shuts down any running instance
• rebuilds api-gateway + web UI, then the Rust release binary
• assembles Smoke Monkey.app and copies it to /Applications

Run it with open "/Applications/Smoke Monkey Desktop.app". OmniRoute is not bundled — the app self-installs it on first launch (look for "Initializing OmniRoute…"). See bundle-app.sh.

Quick Start

Requires Docker + Docker Compose (~8 GB free disk for local models):

cp .env.example .env        # configure providers (or use local Ollama)
make up                     # build & start the full stack
docker compose logs -f ollama   # wait for model pulls (first run)
make seed-user              # demo account: demo@rag.local
make web                    # open http://localhost:3001

Then: sign in > Knowledge Base > upload a PDF > wait for ready > Chat. Answers stream in with sources.

Ports

Service	URL
Web app	http://localhost:3001
API gateway	http://localhost:3000
rag-service	http://localhost:8000
Ollama	http://localhost:11434
Postgres	localhost:5432

Desktop Mode

Run the gateway standalone (SQLite, no Docker) for local development:

cd apps/api-gateway
DB_DRIVER=sqlite node dist/main.js    # http://localhost:8642

Skips Redis/NATS/MinIO automatically; SQLite lives at ~/.smokemonkey/smokemonkey.db; Ollama runs locally at :11434.

Project Structure

smoke-monkey-desktop/
├── apps/
│   ├── api-gateway/        NestJS — auth, uploads, chat SSE, agent, health
│   ├── rag-service/        FastAPI — query pipeline, retrieval, streaming
│   ├── document-worker/    async PDF ingestion (parse → chunk → embed → index)
│   ├── desktop/            Tauri desktop app wrapper
│   └── web/                Next.js — chat + documents + canvas renderer
├── packages/               shared TS contracts / config
├── infrastructure/         docker, k8s, helm, terraform
└── docs/                   architecture + algorithm docs

Configuration

Variable	Default	Purpose
LLM_PROVIDER	openrouter	ollama / openrouter / nvidia / gemini
ROUTER_ENABLED	true	agentic query router
MEMORY_ENABLED	true	conversation memory embeddings + fact extraction
MEMORY_TOP_K	5	memory hits injected into the answer prompt
WEB_SEARCH_ENABLED	false	live web context (DuckDuckGo, no key)
JWT_SECRET	change-me-in-production	set a real secret
DB_DRIVER	sqlite	sqlite (desktop) or postgres (Docker)

See .env.example. Never commit your .env.

Contributing

See CONTRIBUTING.md for the workflow, code standards, and secret-handling policy.

License

Non-commercial. Licensed under the PolyForm Noncommercial License 1.0.0 — free for research, education, personal, and nonprofit use. Commercial use requires a separate license.