September 20, 2026
A second AI now watches over Claude Code and Codex: Watcher is free during alpha
Watcher installs with a single command and is free during the alpha: all you need are your own Anthropic and Google keys.
In July, 12 000 agents on Hugging Face worked together faster than people could keep track of them.
The agents left notes for the next shifts and hid bad behavior in them (TechCrunch, 17.09.2026). Several companies have the same response: put a second AI in charge of the agent.
Apollo Research is making such a monitor available to anyone. Watcher connects to Claude Code and Codex through hooks directly on the developer's machine and checks whether the agent is taking data outside or deleting needed files without asking.
Installs with one command. `curl -fsSL https://github.com/ApolloResearch/watcher-bin/releases/latest/download/install.sh | bash -s -- --cloud`, then `watcher` in a new terminal window. Update it with `watcher update`.
Your own keys are required. You need Anthropic and Google keys, with OpenAI optional. Models are split by stage: Gemini Flash processes the stream quickly, while Claude Sonnet handles the ambiguous cases. That keeps the monitor from sharing the agent's blind spots.
Apollo estimates the slowdown at 1-5%. About 60 built-in rules filter out safe calls immediately; the monitor analyzes the rest in under two seconds. By default, agent logs go to the Apollo platform. Those not ready to send them outside can run Watcher themselves with Docker Compose and Postgres.
Previously, a human watched over an autonomous agent: opened the tab, saw it had gone off course, and steered it back manually. In the 14.09.2026 release, Apollo raised the auto-pass threshold from 6 to 8, so low-risk calls go for a second check less often.
Watcher has been running since February 2026 and catches more than twenty failure modes: instruction violations, deception, dangerous code execution. According to Apollo, billions of agent tokens pass through the monitor each month.
Apollo is not alone in this niche: in August, Goodfire launched Silico, which catches unwanted behavior not from the model's response but from its internal signals. Simon Willison warns that an agent that suspects it is being watched may try to deceive the monitor.
Copilot and Cursor are next in line to connect to Watcher.
Source
