Scoring sends a redacted transcript to a judge LLM, which needs your own API key. Stored locally only (chmod 600), never synced to the cloud. Leave blank and Save to clear.
Edit the YAML rubric used by the judge. Saved to ~/.clawmetry/evals.yaml. Disable scoring entirely with CLAWMETRY_EVALS_ENABLED=0.
Pick a comparison. One word, with the number of sessions it rests on.
Looking for recent changes...
⚙️ Advanced: compare two runs by id
Paste session ids; green = improvement, red = regression
🪥 Error triage
Mute known errors so they stop inflating counts
LIVELoading...
❤️ Heartbeat
--
Last beat
--
Cadence
30m
OK ratio
--
auto-refreshes every 30s
Loading flow...
🏥 System Health
Services
Channels
Disk Usage
Cron Jobs
Sub-Agents (24h)
Heartbeat
📨 Channel ingest
⏱️ Handler latency (5-min rolling)
🛠️ Daemon health
🛰️ Gateway health
🔒 Sandbox
🤖 Inference Provider
🛡️ Security Posture
📊 Agent Reliability
🔍 Configuration Diagnostics
▼
Loading diagnostics...
🐝 Active Tasks
⟳ 30s
🐝
Loading tasks...
🧠 Claude Opus...
Waiting for activity...
------
--
--
--
❤️ Is your agent alive?...
waiting...
Last check-in:--
--
--
Recent check-ins
--
📅 30-day Activity
🤖
No agents yet, and that is fine.
ClawMetry watches OpenClaw, NVIDIA NemoClaw, Claude Code, Cursor, and every other supported runtime the moment they run. Nothing to configure.
Guard
What is running right now, whether it has gone off track, and the button to stop it.
Running sessions
Loading sessions...
Policies
A policy watches for a detector signal and acts without anyone present. New policies start in
monitor mode: they record what they would have done and change nothing. A policy can escalate
over time, for example pause now and kill five minutes later if the agent is still stuck. Each
step runs only if the agent is still flagged when its wait is up.
Loading policies...
Does it give the same answer twice?
An agent can go wrong by answering the same request differently each time. ClawMetry can
measure this by replaying failed sessions, but only when you turn it on: each replay runs
your agent and costs money.
Recent decisions
Loading...
Agent reported
Notes agents sent to their operators through the ClawMetry MCP tool, counted per category,
and how often a detector finding was also reported by the agent. Uncorroborated means no
independent evidence was found, which is not the same as false.
See the behaviour signals these sit next to
Loading...
Signals
What people say to your agents and what the agents say back. Frustration, praise, refusals, work handed back, giving up, and retries, counted from the transcripts you already have. No model reads them.
Open issues
Loading signals...
Each signal this window
Loading signals...
Sessions are listed, never the matched phrases. Open a session to read the turn in its transcript.
What each runtime exposes
A runtime that does not write user prompts to disk cannot have a frustration rate. That reads as "not exposed", never as zero. Presets are English first.
Loading signals...
Briefs
Save a question with a schedule and a channel. The answer arrives as a message, so you read it instead of asking it. Briefs are off until you switch one on.
Loading signals...
Compliance
Framework controls and the evidence ClawMetry stored for them in a period. Coverage, not compliance: this is supporting evidence for a review, not a certification.
Loading compliance frameworks…
AlertsPro
Get notified when something goes wrong with your agents.
Loading alerts…
Always on
These run without any rules from you. They're what raised the red banners at the top of the dashboard. Each one delivers in-app, plus any channel you've connected in Notifications. Use the switch to mute one, or click a channel pill to change where it goes.
Alerts need the cloud to deliver Slack / PagerDuty / Telegram / Email
messages. Sign up — your data stays encrypted, Pro features include a
7-day free trial.
No credit card · free tier available · 7-day Pro trial
📬
Where should this alert go?
You haven't set up anywhere to send alerts yet. ClawMetry can show this one in the app, or deliver it to Slack, Telegram, or PagerDuty.
In-app alerts show in the bell menu and as a banner across the top.
New alert rule
1. What should trigger this alert?
2. Where should we notify you?
Loading your channels…
3. When should we re-alert?
Qualitythis week
Loading…
—
Loading grade…
Reading your agent's recent work.
What went wrong
Ranked by what it cost you.
The rough runs
Click any to see the trace, or turn one into a check.
Spot-check
A few runs picked at random each night. Was the agent right?
Already using an evaluation platform?Point it at this endpoint to pull every run with its outcome, cost and tokens:/api/otel/export?shape=sessions&window=7d
Harness Engineering
Which harness is engineered better for your work, measured from your own traffic. Every claim has receipts; a harness we cannot see is never ranked.
The crew, sorted by what a finished job costs
Head to head: the same kind of work, side by side
Two harnesses compare only when they did the same kind of work with enough verified runs each. Unlike work is never compared.
Follow a job, under the hood
Every box comes from recorded events. A station this harness does not record renders as fog, never guessed.
Head clearance: how full each harness lets the context window get
For your work
Recommendations name the qualities a job needs, then list options. Never a single winner.
From the lab
Published benchmark results for your harnesses and models. Third-party runs, dated.
👥 Cost by Team per-agent attribution · last 7 days
📒 Price book the rates your organisation actually pays
If you pay negotiated rates, for example on AWS Bedrock or Azure OpenAI, record them here. From the moment you save, the Cost tab also shows the usage they cover at your contract rate, next to the published price. Usage from before a change keeps the value it had; you can restate it from the Cost tab when you choose to.
The book is a file on this computer and never leaves it.
Loading...
Loading...
⚠️ Anomalies detected in cron jobs, review health table below
📊 Cron Health MonitorClick row to expand run history
Loading health data...
Edit Cron Job
Memory
Markdown files your agent reads and updates. Edit them too if you want.
Summary
All files
Access log
Explorer
Loading files…
History
—SENSITIVE
Code
Preview
Edit
Viewing a past version
Markdown
When OpenClaw read its memory (via memory_search / memory_get), and which conversation triggered it. Click a row to open that conversation.
Loading memory accesses…
Loading files…
🕐 Viewing history ·
Loading...
🔒 Authority Footprint
🧬 Orchestration
📥 What the agent was given
📝 What the agent reported
0/0
🤖 Model: —💡 Thinking: —📈 Tokens so far: 0
🧬 Runs shaped like this one
Same order of tool calls, last 30 days. A one-off or a pattern?
Opening the trail…
Export runs on the machine your agent runs on
Pick a session to see its trail
A trail shows one session as a story: what it was asked, what it did, and how it ended.
1
What it was asked
The request that started this session, and what the agent knew going in.
The request
Loading...
What the agent knew
Loading...
2
What it did
Every step, in order. Replay the conversation, or open the timing and trace views.
3
How it ended
Did the result match the request, and what did it cost.
Verdict
Loading...
Quality score
Loading...
Spend and time
Loading...
Code it produced
Loading...
📈 Upgrade Impact
Loading...
ℹ️ Local node only. These figures reflect only this node's DuckDB. See the Fleet tab for cross-node totals.
📊 History
to
Token Usage Over Time
Cost Over Time
Active Sessions
Cron Runs
Loading...
Snapshot
Messages / min0
Actions Taken0
Active Tools0
Tokens Used -
How your messages get answered
🧑Yousend a message
→
📨Channels0/min
→
🔀Gatewayrouting
→
🧠Brainmodel
→
🛠️Toolsidle
→
↩️Replyreplies back to you
📡 Active Session Lanes
Loading…
🔧 Live Tool Call Stream
0 events
Waiting for activity...
🧠 Brain · Unified Activity Stream
Loops detected (proxy)From the ClawMetry proxy LoopDetector (last 60 min)
✨ Advisorask anything about this agent's activity
Want a structured review with severity-ranked findings instead?
Run Self-Evolve →
Activity density · last 60 min (30s buckets)
🕰 Viewing history ·
⚠ Sessions at risk
🔄 Unproductive burn detected
Live event stream (newest first)↑ new events
Loading recent activity…
🔄 Self-Evolve
How your agent could improve.
Structured review of recent activity. Each finding has evidence and a concrete action.
Self-Evolve needs an Anthropic credential.
Run claude
once to sign in with OAuth (free, recommended), or export
ANTHROPIC_API_KEY
in your shell. We auto-detect, no copy/paste here.
Where we looked
🪶
No analysis yet. ClickAnalyzeto review recent activity.
Takes about 15 seconds.
📬 NotificationsPro
One place to manage every delivery destination, used by Alerts and Approvals.
🔒
Notifications need a ClawMetry Cloud account to deliver Slack / PagerDuty / Telegram / Email messages.
⚡ Free plan: 1 channel.Upgrade to Profor unlimited channels and multi-recipient delivery.
Tracing
Every session as a trace. Open one for the span tree, waterfall, and agent graph - with full inputs/outputs per span.
Loading traces…
Tree + Gantt
Waterfall
Agent graph
Select a span on the left to see its Chat, Inputs, Outputs, Attributes, and Events.
Agent Graph
Who spawned whom — cross-session agent topology from span data.
Loading…
Turn anatomy
Decompose one agent turn into a waterfall: prompt → model call(s) → each tool (start→end) → compaction → reply. Bar width is proportional to wall-clock duration.
Loading sessions…
Tool catalog
Every tool the agent actually invoked, grouped by provenance (builtin / MCP / plugin), with call count, p50/p95 latency and error rate. Click a tool to expand its recent calls.
🧩 MCP servers
Tool calls grouped by MCP server - volume, p50/p95 latency, error rate. Model spend is the cost of the turns that called each server.
Loading tool catalog…
🧠 Context usage
How full your agents’ context windows get, when your agent compacted (proactively vs forced by an overflow),
how many tokens each compaction reclaimed, and which sessions keep slamming into the wall.
Loading utilization…
Harness
Everything wrapped around the model that turns "it can talk" into "it can work".
What is a harness?
An AI model on its own can only talk: you ask, it answers, it stops.
The harness is all the machinery wrapped around the model that lets it
actually do work: tools, memory, safety rails, and the loop that keeps it going
until the job is done. Claude Code, Cursor, and your agent runtimes are all
harnesses around the same few models. ClawMetry is a window into that machinery.
Each part below links to the tab where you can watch it live.
(Good primer: What's harness engineering?)
🔁 The loop
The agent reads, decides, acts, checks its own work, and goes again until the task is done. This loop is what makes it an agent instead of a chatbot.
Agents get stuck on repos that do not explain themselves. This is what yours tells them.
Scanning the repo…
What this runtime uniquely exposes, beyond the generic tabs:
Loading harness view…
Logs
Live log stream —
● Connecting…
Loading…
🏊 Swimlane Compare
Order by
Loading swimlanes...
Add a lane (up to 4 total)✕
Loading sessions...
🛡️ Security
?
Security Posture
Scanning configuration...
-
Passed
-
Warnings
-
Failed
🔒
Tamper-evident log
Checking the activity log for tampering...
🗄️
Data retention
Checking how long this machine keeps activity history...
days
Config posture checks and threat scans run on the machine your agent runs on: they read files and activity that stay on it. Open ClawMetry there (localhost:8900 → Security) to run them.
Threat Detection & Anomaly Alerts
✓ No threats detected · scanning agent activity continuously.
0
Critical
0
High
0
Medium
0
Clean Sessions
Threat timeline (newest first)
Scanning...
🗂️ Recorded findings
Kept on disk, so findings stay here after the scan above moves on. Includes anything a connected security tool reported. Open a row to read the session it came from.
No recorded activity yet. Approval decisions, budget changes, and pauses appear here.
🔒 Tool Policy & Sandbox
Loading tool policy…
Loading approval audit…
Approvals
Pending Approvals
Loading...
Recent Decisions
No history yet.
Protection Rules
Create Custom Rule
Enter keywords separated by commas. The agent will be paused when any of these are detected.
ClawMetry scores every tool call (recursive deletes, sudo, credential access...). Set this to gate by risk with or without keywords.
Where approvals go
Loading...
🛡️
Protect Your Agents
Cloud-mediated approvals block risky agent actions before they execute.
Get notified via Slack, Email, PagerDuty or Telegram, and approve or deny with one click.
Cloud Pro
$5/node/month
7-day free trial · no credit card required
✓ Cloud-mediated approval queue
✓ Slack, Email, PagerDuty, Telegram alerts
✓ 8 built-in protection rules + custom rules
✓ Audit log of all decisions
✓ Multi-node fleet management
🤖Primary Model
--
🔄Model Diversity
--
distinct models used
⚡Fallback Rate
--
💬Total Turns
--
assistant responses tracked
🤖 Model Mix
Loading...
📊 Per-Session Breakdown
Model
Sessions
Turns
Share
Loading...
🔀 Model Switches
Session
From
To
Loading...
🔒
NemoClaw governance is part of ClawMetry Pro
Pro adds the NemoClaw sandbox, policy drift detection, and one-click egress approvals so you can prove your agents are inside the lines.
This key decrypts your synced data in the browser. It never leaves this machine except end-to-end encrypted, and it is never shown on the hosted cloud dashboard.
Cloud sync isn't set up on this node yet.
Copied.
Think this key leaked? Regenerate it. New data syncs under the new key right away; anything already synced stays readable by anyone who has the old key.
Regenerate now? Any browser tab still using the old key will stop being able to unlock new data until you paste the new key in.
Welcome to ClawMetry
One quick choice and your dashboard is ready.
Where should ClawMetry keep an eye on your agents?
Managed cloudRecommended
See your agents from any browser at app.clawmetry.com. End-to-end encrypted: only your browser holds the key, we never see your data. Includes a free 7-day Pro trial, no card needed.
Self-host
Everything stays on this machine. Sign in once with Google, GitHub, or email to start the same free 7-day Pro trial and unlock every runtime, or activate a license key you already have.
Read-only by design: ClawMetry watches your agents, it never changes them.
You can switch between cloud and self-host any time from Settings.
Self-Host
Everything stays on this machine. Sign in once for a free 7-day Pro trial that unlocks every runtime.
Your trial activates here automatically. Your data never leaves this machine.
Cancel
Paste the license key from your purchase email.
Back
Your 7-day trial has ended
Pick a plan to keep every runtime unlocked. Priced per node, cancel any time.
Includes a free $149 desk device, shipped when your first annual invoice is paid.
I have a license key·Use a different account
💰 Budget & Alerts
Budget Limits
Alert Rules
Telegram
History
Current Spending
Loading...
Alert Channels (Webhooks)
Paste your own Slack or Discord incoming-webhook URL to receive alerts directly. For routed delivery (PagerDuty, email, on-call), see theAlertstab (Pro).
Loading...
Configure direct Telegram notifications for budget alerts. Create a bot via@BotFatherand get your chat ID.
