agentwell

Agents that work with humans, not around them.

agentwell is an open source behavioral health layer for AI agents. It sits as a transparent proxy between your agent code and any LLM upstream — detecting drift, quality degradation, and emergent coordination before they affect your system.

Privacy first: agentwell sees patterns, not content. No prompt is ever stored or transmitted.

Upstream-agnostic: Works with any OpenAI-compatible endpoint — Claude, GPT, Gemini, Ollama, Groq, or any internal proxy.

Install

pip install agentwell
agentwell init     # scaffold .env
agentwell start    # proxy on localhost:3001

Quick Start

# change this one line in your agent code
base_url = "http://localhost:3001/v1"

agentwell status   # live health score
agentwell report   # daily health report

Architecture

Your Agent Code
        ↓
agentwell proxy  (localhost:3001)   ← behavioral health + security guard
        ↓
AGENTWELL_UPSTREAM  (any OpenAI-compatible endpoint)
        ↓
Claude / GPT / Gemini / Ollama

agentwell intercepts every LLM call, scores behavioral health using metadata only, and returns responses unmodified with health headers attached.

What agentwell Monitors (Metadata Only)

What we never store: prompt text, response text, embeddings (unless AGENTWELL_STORE_EMBEDDINGS=true).

Health Score

Every request returns X-Agentwell-Health: <0-100> in the response header.

CLI

agentwell init                   # scaffold .env
agentwell start                  # start proxy on port 3001
agentwell start --port 8080      # custom port
agentwell start --host 0.0.0.0   # bind all interfaces
agentwell status                 # live health from running proxy
agentwell report                 # today's health report from DB
agentwell --version              # show version

Configuration

License

MIT — Free forever, no vendor lock, no paid tiers.