governance toolkit for AI agents

A provider-agnostic governance kernel with adapters for LangChain, Claude, and your own stack. Context compression · tool allow/deny policy · token budgets · opt-in telemetry.

Use it as a LangChain 1.0 middleware, a Claude adapter, a standalone SDK, or a terminal command. Same governance model across all four.

$ pip install axor-langchain # LangChain 1.0 middleware $ pip install axor-core # provider-agnostic kernel $ pip install axor-cli # governed sessions in your terminal
View on GitHub →

Benchmark

3-node research pipeline · 8 turns · real API calls
−26.2% input tokens
claude-sonnet-4-6 · benchmark/live_graph.py · reproducible across runs

How it works

  agent turn N
      │
      ▼
  ┌─────────────┐    ┌────────────────────┐    ┌─────┐
  │  messages   │───▶│  axor middleware   │───▶│ LLM │
  └─────────────┘    │  · compress        │    └─────┘
                     │  · filter tools    │
                     │  · enforce budget  │
                     └────────────────────┘

Packages

axor-langchain
LC 1.0 AgentMiddleware — compression, tool governance, budgets.
pypi · repo
axor-core
Provider-agnostic governance kernel. Policy engine, context shaping, federated child spawning. Zero runtime deps.
pypi · repo
axor-cli
axor command — governed agent sessions from the terminal. Claude / OpenAI adapters.
pypi · repo
axor-telemetry
Anonymous opt-in pipeline. MinHash fingerprint only, no raw input ever.
pypi · repo
axor-claude
Claude Code adapter for axor-core.
repo
axor-memory-sqlite
SQLite-backed cross-session memory provider.
repo

Demo

axor-langchain asciinema demo — install, middleware, turn-by-turn compression output

60 seconds · install · plug in the middleware · watch compression cap token growth.

Community