Activity without context
An agent calls tools and changes files. You need the session behind those actions to understand what happened.
AI endpoint observability
Your AI agents.
In plain sight.
Your team works with Claude Code, Codex, Cursor and OpenClaw. See what those agents do, understand their costs, and catch the runs that need your attention.
Local-first observation. Optional controls. Your existing workflow.
From activity to understandingIllustrative example
Claude CodeCodexOpenClaw
ClawMetrySessions, tools and usage
An agent is repeating the same action
Open the session. Follow the evidence. Decide what happens next.
Built for the agents your team already uses
Your agents are busy.
Do you know what they are doing?
Work spreads across terminals, editors and machines. A running process tells you very little about the work happening inside it.
Cost without progress
Repeated failures and loops can keep consuming tokens long after useful work has stopped.
Autonomy without oversight
When behavior looks wrong, your team needs inspectable evidence and a clear way to respond.
See the work.
Understand it. Take control.
One place to follow your agents, investigate problems and decide when to step in.
See
A shared view of your agent activity.
Bring supported runtimes into one dashboard. Follow sessions, inspect tool calls and compare usage across connected machines without stitching together separate histories.
Explore runtime coverageAgent activityIllustrative example
Claude CodeWorking through a code change
ActiveCodexRepeating a failed command
Needs reviewOpenClawRecent session available to inspect
CompletedUnderstand
Find the problem.
Follow the evidence.
Investigate loops, repeated tool failures and unusual behavior. Open the sequence behind a finding, see the affected session, and review cost estimates where the runtime provides usage data.
See what Guard checksA finding you can investigateIllustrative example
A tool call failsThe result is recorded in the session.
The agent keeps repeating itThe same action appears again in the sequence.
Guard flags the patternReview the calls and results behind the finding.
Intervene
Choose when
to step in.
Start with monitoring. Pause or stop a supported session when you need to, or opt into policies that respond to findings. Supported tool integrations can hold an action for approval before it runs.
Check controls and their limitsA deliberate responseIllustrative example
- When
- A stuck loop is detected
- Start with
- Monitor and record
- After enabling enforcement
- Pause a supported session
Automatic enforcement is off by default. Controls depend on the runtime and session.
Questions,
answered.
Which agents can I observe?
ClawMetry supports runtimes including Claude Code, Codex, Cursor, OpenClaw and Hermes. Available history, usage and controls vary by runtime. The integrations page explains the scope of each connection.
Will ClawMetry change how my agents work?
Observation is read-only by default. Manual controls require your action. Automatic enforcement and pre-tool approval integrations require separate setup and support.
Does it prevent every harmful action?
No. Detection uses the activity ClawMetry can observe. A finding after an action is different from a gate before it. Pre-tool approval is available through supported integrations; it is not universal endpoint protection.
Can I see exactly what every agent costs?
Cost visibility depends on the runtime. Some provide usage that ClawMetry can value using model rates; others do not expose cost data. Estimates and unavailable values are identified rather than presented as a provider invoice.
What will we cover in the demo?
A 30-minute walkthrough of agent activity, session evidence, available cost data and supported controls. We will discuss the runtimes your team uses and what a rollout would look like.
Put your agent activity
in plain sight.
See ClawMetry in action and map it to the way your team works.