Free version self-serve
Free source (opens in a new tab)
N/a
For: Trying monitoring
Limits
- Claude Code and Codex only
Includes
- CLI install
- Hosted by Apollo
(open full size in a new tab)
(open full size in a new tab)
(open full size in a new tab)
(open full size in a new tab)
(open full size in a new tab)
Our take: Consider Watcher if your engineers run Claude Code or Codex and your security team wants central policies that block risky actions such as unreviewed terraform apply before they execute.
Main limitation: Self-serve install covers only Claude Code and Codex; other harnesses and self-hosting go through a sales call.
Is Watcher (by Apollo Research) right for you?
Consider Watcher if your engineers run Claude Code or Codex and your security team wants central policies that block risky actions such as unreviewed terraform apply before they execute. Based on the vendor's documentation, not on a hands-on test.
Best for: Engineering teams running coding agents
Also a good fit for
- Security teams
Consider another tool if self-serve install is for Claude Code and Codex only
Or if
- Hosted install connects to a monitoring platform in Western Europe
- Enterprise requires booking a call
Watcher (by Apollo Research) pricing and usage limits
Prices checked Sep 29, 2026 on the official US pricing page. Prices and limits change often.
Free access and trial conditions
A free version is described in the announcement
Paid plans: what changes as you upgrade
Enterprise sales
Custom pricing source (opens in a new tab)
Custom
For: Organizations
Includes
- 50+ agent harnesses and gateways
- Self-hosting
1 more
- Central policy rollout
Usage caps, credits and extra costs
Enterprise priced via a call
What is Watcher (by Apollo Research), and how does it work?
Watcher by Apollo Research monitors coding-agent sessions, records what agents do, grades activity with LLM graders, can intervene in real time, and provides a dashboard. The self-serve version integrates with Claude Code and Codex via a command-line client; an enterprise version is described as supporting more harnesses. It can be hosted by Apollo or self-hosted. Its Watcher Live component checks actions before they run and blocks risky ones; Watcher Analyze reviews sessions. The announcement describes a free version and an Enterprise tier via a sales call.
What you provide and what you get back
You provide
- Agent actions and sessions from Claude Code or Codex
- Organization policies and risk thresholds
- Rollout via MDM
You get back
- Real-time allow or block decisions on actions
- Recorded sessions with graded findings
- Dashboard with recommended policy updates
Where you can use it
- Platforms
- Web, API
- Account
- Required
- Supported agents (self-serve)
- Claude Code, Codex
- Supported agents (enterprise)
- 50+ harnesses and gateways
- Hosting
- Apollo-managed (Western Europe) or self-hosted
- Stated critical-failure recall
- 100% on a synthetic attack dataset
- False positives
- <1% on benign tool calls
- Overhead
- 3-5% cost, 6-8% latency
Where Watcher (by Apollo Research) is most useful
Scenarios are worked out by our editors from the documentation; the actual result depends on your input and plan.
Blocking risky agent actions before they run
- Who
- Security teams governing coding agents
- You provide
- Agent tool call such as terraform apply -auto-approve
- Expected result
- Action blocked when it crosses the risk threshold
- Plan needed
- Free version or Enterprise
- Limitation
- Stated results come from a synthetic attack dataset source (opens in a new tab)
Reviewing coding-agent sessions after the fact
- Who
- Engineering leads investigating incidents
- You provide
- Recorded sessions
- Expected result
- Graded findings with a recommended action
- Plan needed
- Free version or Enterprise
- Limitation
- LLM graders can be wrong; check flagged items source (opens in a new tab)
Rolling out central agent policies via MDM
- Who
- IT and platform teams
- You provide
- Policy definitions pushed to devices
- Expected result
- Protected devices across the engineering team
- Plan needed
- Enterprise (book a call)
- Limitation
- Requires organization-level deployment source (opens in a new tab)
Monitoring several agent harnesses at once
- Who
- Enterprises using many agents
- You provide
- Agent gateways and harnesses
- Expected result
- One view across 50+ supported harnesses
- Plan needed
- Enterprise
- Limitation
- Self-serve supports only Claude Code and Codex source (opens in a new tab)
What Watcher (by Apollo Research) does, and where it stops
Real-time action blocking
Checks each agent action before it runs and blocks those over a risk threshold Stops destructive commands rather than reporting them later
Session recording and trajectories
Records what agents did in each session Gives an audit trail for investigation
LLM graders
Uses LLM graders to rate activity and surface findings Prioritizes which sessions to read
Central policy rollout
Defines rules once and deploys them across devices through MDM Consistent policy for every developer
Self-hosting
Runs Watcher in your own infrastructure Keeps agent data inside your environment
Watcher (by Apollo Research) pros and cons
Reasons to choose Watcher (by Apollo Research)
- Supports Claude Code and Codex out of the box Two common agents can be monitored with one install command documented source (opens in a new tab)
- Self-hosting available Sensitive teams can keep monitoring in-house documented source (opens in a new tab)
- Docs are published as markdown with llms.txt Agents can read the setup documentation directly documented source (opens in a new tab)
- A free version is stated to exist Teams can try it without a contract documented source (opens in a new tab)
- Monitoring models configurable independently of the agent The checker is not the same model as the agent documented source (opens in a new tab)
Check before paying for Watcher (by Apollo Research)
- Self-serve install is for Claude Code and Codex only Other agents need Enterprise documented source (opens in a new tab)
- Hosted install connects to a monitoring platform in Western Europe Data location matters for compliance documented source (opens in a new tab)
- Enterprise requires booking a call No self-serve price documented source (opens in a new tab)
- Adds an estimated 3-5% cost and 6-8% latency Agents run slightly slower and costlier documented source (opens in a new tab)
- Detection figures come from a synthetic dataset Real-world recall may differ documented source (opens in a new tab)
How to use Watcher (by Apollo Research) for monitoring Claude Code sessions
Set up Watcher (by Apollo Research)
- Open the install section on watcher.apolloresearch.ai
- Run the curl install command with the --cloud flag
- Restart or open a new terminal
- Run watcher and follow the setup documentation to connect Claude Code or Codex
Example: monitoring Claude Code sessions
- Start a Claude Code session with Watcher active
- Ask the agent to run a risky command such as terraform apply -auto-approve
- Observe the action blocked by Watcher
- Open the dashboard to review the session
Review and export the result
- Confirm the blocked action appears in the session record
- Review the recommended action and update the rule if needed
Watcher (by Apollo Research) API
API access and separate billing
Watcher (by Apollo Research) has no public API in its official documentation. Seed lists docs site; contents not in fetched pages API documentation ↗
Privacy, data use and commercial rights
Summarized from the vendor's current policies; this is not legal advice and a vendor's badge is not an independent audit.
How your inputs and outputs are handled
Agent actions and sessions are sent to Apollo's managed platform in Western Europe for the hosted install
Controls for business users
- Self-hosting option
- Configurable monitoring models
Best Watcher (by Apollo Research) alternatives for different needs
"Better suited" is based on documented features and pricing, not on a head-to-head test.
Claude: better suited to
GitHub Copilot: better suited to
| Alternative | When to choose it | Pricing | Limitation | |
|---|---|---|---|---|
| Claude | — | — | ||
| GitHub Copilot | — | — | Compare |
Explore Watcher (by Apollo Research) further
Related to Watcher (by Apollo Research)
Watcher (by Apollo Research) user reviews
No user reviews yet.
Used Watcher (by Apollo Research)? Be the first to share your experience.
Write a review
Watcher (by Apollo Research) FAQ
Which agents does the self-serve Watcher support?
Claude Code and Codex; the enterprise version supports 50+ agent harnesses and gateways. source (opens in a new tab)
Is Watcher cloud-hosted or self-hosted?
The page's install command connects to Apollo's managed platform in Western Europe; you can also self-host by booking a call. source (opens in a new tab)
How much overhead does Watcher Live add?
Estimated 3-5% added cost and 6-8% added latency for the default configuration. source (opens in a new tab)
How is Watcher different from the agent's own auto-mode?
It gives security teams centrally managed policies and visibility across agents, with monitoring models configurable independently of the agent. source (opens in a new tab)
How do I install it?
Run the curl install.sh command from the release page with the --cloud flag, open a new terminal and run watcher. source (opens in a new tab)
Sources and review methodology
Official sources and verification dates
- Watcher documentation homepage · checked Sep 29, 2026
- Announcing Watcher Live homepage · checked Sep 29, 2026
- Watcher home homepage · checked Oct 1, 2026
Not verified: Pricing and free version limits; Privacy policy and retention; Training use; Contents of docs site.
How we documented and maintain this page
Every fact about Watcher (by Apollo Research) above comes from the 3 official pages listed here, last compared on Sep 29, 2026; we have not run Watcher (by Apollo Research) ourselves, so it has no editorial score. How we check and re-check pages.
Article updated Oct 10, 2026; prices checked Sep 29, 2026. Outbound links to vendors are plain links; we do not earn commissions. How we review · Report an error