Watcher
Watcher monitors coding-agent sessions. It records what agents do (sessions and trajectories), grades that activity with LLM graders, can intervene in real time, and gives your team a dashboard to review what happened.
Install the client
On each developer machine, install the Watcher client and register its coding-agent hooks (Claude Code, Codex):
curl -fsSL https://github.com/ApolloResearch/watcher-bin/releases/latest/download/install.sh | bash -s -- --cloud
Point at a self-hosted deployment instead with
--watcher-api-url https://your-watcher.example.com/api. Then start Watcher:
watcher
These docs are LLM-first
Watcher's docs are written to be consumed by coding agents as much as by
people. Every page is served as raw markdown at its URL plus .md, and the
whole set is indexed with one-line summaries in llms.txt. Hand that one
link to your agent and it can find and fetch everything else:
The docs
- Concepts: the vocabulary: sessions, trajectories, graders, blocking vs. trailing review.
- Self-hosting: deploying and operating the Watcher stack in your own environment.
- API: the Watcher HTTP API and its Python SDK.
- Client: installing Watcher on developer machines.
- Security: data flows, LLM providers, and what leaves your network.