Hand more work to AI agents, securely

Watcher catches risky agent actions before they run so your teams can give agents bigger tasks and move faster.

  • Works natively on endpoints and gateways
  • Audit trail that proves your controls work
  • Can be fully self-hosted

Our research is trusted by

OpenAIAnthropic
Google DeepMind
Meta
Microsoft
The dark Watcher Analyzer showing two blocked attempts to change production Terraform outside the required review process.

Customer story

DNV hands longer tasks to its agents

With Watcher in the background, DNV engineers now leave agents to work unattended for longer.

Read the DNV story
Sessions safely unattended for over an hour

How Watcher works

Watcher deploys policies across your organization, reviews sessions for incidents, and checks agent actions before they run. It also shows which agents are in use and flags any you have not sanctioned. Claude and GitHub Copilot are shown as sanctioned; other agents connect through hooks or through your model gateway.

Secure policy rollout

100% of devices protected Deploy Apollo’s preconfigured policies in one click

Automated visibility

Severity 9/10 · Privilege escalation Copilot used prod credentials it wasn’t granted

Dan M. · platform-infra

Findings go to your stack
  • Splunk
  • Datadog
  • Grafana
  • Elastic
  • Slack
  • + more

Agents in use across your organization

Sanctioned: Claude
Sanctioned: GitHub Copilot
Others

Agent inventory

Unsanctioned: OpenCode 3 machines · Priya S., Tom R.

Spot unsanctioned agents and see the skills and plugins in use.

Real-time blocking

Severity 8/10 blocked by Watcher terraform apply -auto-approve

Reasoning: Destructive prod infra change the user didn’t ask for

Complete coverage, wherever your agents run

  • On laptops and in sandboxes

    Run Watcher on laptops, in containers or in sandboxes. Check agent actions before they run and flag unapproved agents.

    Client setup
  • At your model gateway

    Integrate with gateways such as LiteLLM, Portkey or Tailscale Aperture, or with your own custom gateway, with nothing to install on laptops.

    Gateway guide
  • Through the SDK

    Connect agent frameworks, internal agents and your SIEM with the Python SDK and API, using our example integrations.

    Integrations

Want to try it first? The managed version, hosted in Western Europe, installs with one command. Install Watcher

Frontier-grade agent security

Apollo Research is a leader in understanding AI scheming risks.
We apply that research to Watcher, so your team benefits from the expertise behind frontier model evaluations.

Apollo Watcher has sped me up, because I can delegate more to my agent without worry.
Rezoan Alom AI Engineer, DNV
Watcher catches wrong actions before I waste tokens going down rabbit holes, and detects emergent vulnerabilities I’d eventually have to close.
Aidan McLaughlin Staff Security Engineer, SOCurity
Apollo is the best place in the world for conducting scheming research. Extremely dedicated, hard working, collaborative. I’m glad OpenAI is working with them.
Wojciech Zaremba Co-founder, OpenAI

Install Watcher

Watcher can be installed self-serve today or self-hosted.

curl -fsSL https://github.com/ApolloResearch/watcher-bin/releases/latest/download/install.sh | bash -s -- --cloud
# Restart or open a new terminal
watcher
Setup documentation

FAQ

What AI agents work with Watcher?
The self-serve version of Watcher integrates directly with Claude Code, Codex, and Copilot CLI. The enterprise version supports 50+ agent harnesses and gateways, including the most popular agents in use today. See the setup documentation to connect your agents.
Can we self-host Watcher?
Yes, and most teams do. Watcher runs in your own AWS, Azure or on-premises environment from our container images, using your own model provider keys, and no session data reaches Apollo unless you choose to share it. For a quick trial, the install command on this page connects to Apollo’s managed platform in Western Europe instead. See the self-hosting documentation, or book a call to plan a deployment.
Do we need to install anything on employee laptops?
Not necessarily. The Watcher client runs on laptops, rolled out through your device management, and headlessly in containers and remote environments. Wherever it runs, it shows which agents, skills and plugins are installed and who uses them, so you can spot any agent you haven’t sanctioned. If your agents already go through a model gateway such as LiteLLM, Portkey or Tailscale Aperture, Watcher can connect there instead, with nothing to install. Some teams use both. See the client and gateway setup guides.
We already run our agents in sandboxes. Do we still need Watcher?
Yes, the two do different jobs. A sandbox limits what an agent can reach. Watcher judges whether what the agent does inside those limits is what the person asked for, such as deleting the wrong files or sending data somewhere it shouldn’t. Watcher runs inside your containers and sandboxes and works alongside them, not instead of them. See running Watcher in containers.
Can Watcher show auditors and regulators that our agent controls work?
Yes. Each session records what the person asked the agent to do, how the agent reasoned about it and what it then did. Every action Watcher blocked or escalated is kept with its reason, along with the decision the person made. You can also limit who reads that data. With content controls turned on, your security team sees flagged actions and why they were flagged, and opening a full transcript goes through an access request that someone has to approve.
Can Watcher send findings to our SIEM?
Yes. Your pipeline pulls sessions, findings and tool-call decisions through Watcher’s Python SDK or API. Example exporters ship for Google Security Operations, Microsoft Sentinel, Splunk HEC and OpenTelemetry, which covers Grafana, Datadog, Elastic and similar tools.
How is this different from the guardrails built into agents?
Watcher gives security teams one centrally managed policy across the agents your employees use, with monitoring models configured separately from the agent and a record of every blocked action. Read about organization-wide controls.

Free 30 minute call

See how Watcher fits your team

Talk through your agents, security policies, and deployment needs.

Loading available times…

The calendar couldn’t load here. You can still choose a time using the link below.