Numen

Why it's ranked here

The case rests on the join. Per-tool integrations give an agent separate silos; this one hands over a single map where the same person across three apps is already one person and the ranking is already computed. It ships under the AGPL, talks to any OpenAI-compatible model including a local one, and refuses to start in production with shipped default passwords. Against that, the project declares itself beta, its own architecture notes still describe an earlier three-connector scope, and the upgrade notes warn that database migrations cannot be reversed. A serious self-hosted option for teams willing to read the code, not a drop-in product.

What's good

The hardening is specific. The evaluation setup binds Postgres, Redis and the graph database to loopback only, with comments saying why. Production startup fails if a secret is still the placeholder, if the developer header login is on, if webhook signature checks are off, or if either database password is the compose default. A seeded demo organisation needs no model key and no OAuth app, so the graph can be explored from a coding assistant before anything real is connected. Outside production, missing secrets are generated and cached so the documented quick start actually boots.

Tradeoffs

Setup is heavy. Every connector needs its own OAuth app in that vendor's console, and web sign-in needs Google OAuth. Without a public URL for webhooks, syncing falls back to polling every 300 seconds. The default sign-up mode admits anyone at a domain that already has an org, which the README itself calls usually wrong for a company. Six services idle at roughly 340 MB of RAM, the graph database holds everything in memory, and backups must capture two volumes together. The MCP library is pinned below version 2 until the server is migrated. The AGPL obliges anyone serving a modified version over a network to publish its source.

How to use it well

Start with the demo org and point Claude Code or Cursor at it to see whether graph-level answers beat your existing per-tool integrations. Then connect one source, GitHub or Linear, before adding others. Switch sign-up to domain mode with an allowlist before exposing the instance, and set the app URL to the exact hostname people use or agent connections fail with a 421 error. It fits an engineering team that wants cross-tool triage kept in-house. Observability and product analytics sources appear only on the roadmap, so it does not cover incident or adoption signals yet.

Observed

License

AGPL-3.0-only (LICENSE, pyproject.toml)

Language

Python 3.12 backend (FastAPI); React and TypeScript frontend

Packaging

Docker Compose, separate evaluation and production files; numen-mcp console script for stdio

Interfaces

REST API, OpenAI-compatible chat completions endpoint, MCP over HTTP and stdio, Slack, email briefings, web app

Datastores

PostgreSQL with pgvector, Redis, FalkorDB

Model backend

Any OpenAI-compatible endpoint, including local Ollama or vLLM (.env.example)

Declared maturity

Development Status 4, Beta (pyproject.toml)

Default sign-up mode

open: any email at a domain that already has an org (src/config.py)