Security and observability layer for coding agents that records sessions, grades trajectories, blocks risky actions, enforces organization-wide policies and provides audit/incident review.
Watcher is Apollo Research's security and observability layer for coding agents. It records agent sessions and trajectories, grades activity with monitoring models, can block risky actions before they execute, and gives teams a dashboard for reviewing sessions, incidents and policy decisions.12 The current self-serve client hooks directly into Claude Code and Codex; older product material describing Cursor and Gemini CLI as priorities should be read as roadmap rather than current support.34
The product combines several control layers: organization-wide rules, real-time blocking review, trailing analysis and audit history. Teams can use Apollo's hosted cloud or point the client at a self-hosted Watcher deployment, and current documentation exposes an HTTP API and Python SDK for integration.2
Apollo currently describes hosted Watcher as free to sign up for and free to use. The self-serve installer connects local coding-agent hooks to Apollo's cloud without requiring users to provision their own backend.3 Teams that need self-hosting or managed deployment can contact Apollo, but public commercial pricing for those configurations is not listed.
The client currently runs on Apple-silicon macOS and x86_64 Linux. A web interface provides Live/Analyzer views, while self-hosted deployments can place the backend inside an organization's own infrastructure.42
Apollo's credibility in agent monitoring is a meaningful strength, but product claims still need to be classified correctly. The organization publishes research on coding-agent monitoring, calibration, prompt design and red-teaming, and explicitly describes Watcher as the production route for that work.5 The self-serve launch says its monitoring pipeline was tuned on tens of thousands of coding-agent trajectories and more than 40 failure modes and backtested against internal golden datasets.3 Those are useful first-party validation signals, not independent benchmarks.
The data path also matters. In the default hosted configuration the local client records coding-agent sessions and sends them to the configured Watcher backend for grading and history; organizations with stricter data-control requirements can self-host.24 Security teams therefore need to evaluate the monitoring service's own deployment, retention and provider configuration alongside the risks it is intended to reduce.
Independent exact-product evidence remains thin. Current external coverage does not provide a substantial Watcher-specific verified-user rating, comparative product benchmark or independent production study measuring false positives, missed incidents or engineering overhead.
Watcher is strongest for teams already granting coding agents meaningful repository or system permissions and wanting a consistent policy layer across developers. Centralized rules, blocking review and trajectory analysis are more relevant to security-conscious engineering organizations than to an individual using an agent only for low-risk local experiments.12
It is a weaker fit where Claude Code/Codex are not the primary agents, where session data cannot leave the local environment and self-hosting is not practical, or where the organization expects mature independent effectiveness benchmarks before deploying another security control.4
The clearest strengths are research depth, a concrete control model and deployment flexibility. Watcher goes beyond passive logging by combining enforceable policy, real-time blocking and post-session analysis, and its self-hosting/API options make it more adaptable to team infrastructure than a fixed SaaS dashboard.25
The
evidence and maturity. Current support is concentrated on Claude Code and Codex, public enterprise pricing is opaque, and the most detailed effectiveness claims come from Apollo's own datasets and backtests.34 Watcher is therefore a credible security proposition, but its operational accuracy and cost-benefit profile still need stronger independent validation.
Products matched by shared tasks, audience, workflow, product type and capabilities.