Today
Loading task outcomes...
show details
⚡ Agent efficiency playbook
Use measured evidence to get more useful work from your agents.
1. Parallelize independent work
Activity shows active sessions, throughput, and unproductive runs.
Open Activity →
2. Continue work in the same session
Context usage shows prompt reuse, compactions, and what was measured.
Open Context usage →
3. Let unattended work be watched
Guard and Sessions show what is still running, waiting, or needs review.
Open Guard →
4. Route by outcome and cost
Models, Quality, and Cost together show whether a routing change helped.
Review routing →
🎯 How independent is your agent?
Unknown
Unknown
Last 7 days
LLM judge quality
Unknown
Task outcomes count whether a run finished successfully. The judge score checks how the work was done and needs a configured judge. Open Quality for the evidence →
🔬 Evaluators on your agent
Open Evals →
⚖️ Did the change help?
Select a comparison. The result includes the number of sessions used.
Looking for recent changes...
⚙️ Advanced: compare two runs by id
Paste the session IDs. Green indicates improvement. Red indicates a worse result.
🪥 Error triage
Mute expected errors to exclude them from the counts.
Recurring errors
See which errors repeat, then open a recorded example.
Loading flow...
🏥 System Health
Services
Channels
Disk Usage
Cron Jobs
Sub-Agents (24h)
Heartbeat
🔍 Configuration Diagnostics
▼
Loading diagnostics...
🐝 Active Tasks
⟳ 30s
Loading tasks...
🧠 Claude Opus
...
Waiting for activity...
❤️ Is your agent alive?
...
waiting...
Last check-in: Unknown
Unknown
Unknown
Recent check-ins
Unknown
Your machine, explained.
Node activity Reading this node
Waiting for a recorded observation
This view covers the whole node, including human, agent and unattributed activity. A runtime filter does not limit it.
Reading recorded observations from your node.
Starting live measurementsCPU and memory update while you watch. Other observations update separately.
What was observed
Choose an observation below, or replay the sequenceRecorded observations will appear here after collection.
Follow the sequence
Recorded observations and changes0 events
No recorded events are available.
ClawMetry resource usage
Waiting for process identificationUsage for each identified process. CPU: 100% means one core. Memory is resident memory (RSS). Shared memory can appear in multiple processes.
ClawMetry resource observations are not available.
Observed processes
Observation times shown belowNames identify programs, not who started them.
No retained processes.
Process observations are not available.
Observed listening ports
Observation times shown belowA listening port accepts connections. It does not prove Internet access.
Listener observations are not available.
Choose what ClawMetry observes
These settings change collection on this node. They do not change what a process can do.
What we can see
Your event reviews
Historical context onlyMarking a past change as expected does not authorize a future action. Original observations stay in the event history.
Changes worth a closer look
Review new listeners and selected setting changes. An unknown actor is not evidence of an attack.
Observation limits
Periodic collection can miss short processes and temporary connections. A listening port does not prove external access. File fingerprints show a difference, not who made it. This is not a complete record of execution.
YOUR AGENT SETUP
See what is shaping your agents.
Rules, skills, commands, sub-agents, and hooks are spread across different harnesses. ClawMetry gathers the files it can find into one quiet, reviewable view.
Harnesses foundLoadingon this machine
Setup filesLoadingrules and assets
ProjectsLoadingproject-scoped roots
GlobalLoadingglobal roots
Loading setup map…
GUIDANCE LOOP
Turn repeated friction into better defaults
Every card below is something you said to an agent: a correction, a preference, or the same frustration again. Press Investigate and the Assistant reads the sessions behind it. It reports what went wrong and which file to change.
Read-only candidatesNo files change from this view. Investigating only reads your sessions.
CandidatesLoadingsignals to review
ConversationsLoadingwith evidence
ProjectsLoadingacross your workspace
Window30dlocal history
These are evidence-backed candidates. Investigate one to find the cause, then decide whether a rule, skill, or instruction should change.
Reading recent conversations…
Guard
See what is being checked. Decide what happens next.
Reading check settings from your node...
These settings apply to all runtimes on this node. Checks find evidence. They do not block an action on their own.
What needs your attention?
Approvals ask for a decision. Findings explain what a check noticed.
Findings on running sessions
Loading sessions...
Choose how to respond
Keep watching, notify someone, ask for approval, or take an automatic action.
Policies
A policy can act automatically when a detector reports a condition. New policies start in monitor mode. They record the proposed action without changing the agent. An enabled policy can escalate its actions over time. For example, it can pause an agent, then kill it five minutes later if the condition continues. Each step requires the condition to remain active when the waiting period ends.
Loading policies...
Does it give the same answer twice?
An agent can give different answers to the same request. Enable replay to measure this variation in failed sessions. Each replay runs your agent and can incur costs.
Finding history
Recorded loops and repeated tool failures, including recovery and acknowledgement.
Loading finding history...
What happened
Review automatic responses and agent reports. For permission decisions and delivered notifications, open their history.
Recent decisions
Loading...
What your agents load
This view shows components read by Claude Code, Codex, Cursor, Gemini CLI, opencode, and OpenClaw. It includes MCP servers, skills, plugins, instruction files, and hook files. This view combines these runtimes. New or changed components flag the active sessions that load them. ClawMetry reads these files without starting a component. It does not inspect MCP tool descriptions.
Loading...
Agent reported
Notes that agents sent through the ClawMetry MCP tool are counted by category. This view also shows how often an agent reported the same condition as a detector. Uncorroborated means that no independent evidence was found. It does not mean that the report is false.
See related behavior signalsLoading...
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.
Loading signals...
Each signal this window
Loading signals...
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…
Alerts Pro
Alerts tell you what happened. Guard checks find evidence. Approval requests ask you to make a decision.
Notification rules and monitors
Loading alerts…
Always on
These monitors run without custom rules. They produce the red banners at the top of the dashboard. Alerts appear here and in channels connected through Notifications. Use a switch to mute a monitor. Select a channel to change where its alerts go.
Loading monitors…
Recent alert history
Recent notifications on this node. Repeated notifications are grouped. Open Guard to change detection settings.
Loading history…
Delivery channels: Slack · Email · PagerDuty · Telegram
shared with Approvals
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
Compare agent runtimes using measurements from your activity. Each comparison shows its evidence. Runtimes without observed data receive no ranking.
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.
📊 Today
--
📅 This Week
--
📆 This Month
--
📊 Token Usage (14 days)
Loading...
💰 Cost Breakdown
| Loading... |
🏆 Top Sessions by Cost which session burned the budget · top 20
| Loading... |
🧩 Cost By Plugin / Skill
Loading...
💸 Top Sessions by Cost
Alert threshold:
$
per session
Loading...
💱 Cost Comparison what same workload costs elsewhere · 30 days
💡 Spend Optimization model downgrade suggestions · 30 days
🔮 Trace Clusters auto-group sessions by behavior pattern
Loading...
📅 Activity Heatmap hourly usage intensity
Loading...
🎯 Skill Cost Leaderboard heuristic attribution, 30s window
Loading...
What your agents delivered
Your model bill says what the agents cost. This says what that money bought: the pull
requests they worked on, and what each one took. Spend is split between the pull requests
a session touched, never copied onto each.
Loading delivery…
📒 Price book the rates your organisation actually pays
If you pay negotiated rates, record them here. Examples include AWS Bedrock and Azure OpenAI rates. After you save, the Cost tab shows applicable usage at your contract rate beside the published price. Earlier usage keeps its previous value. Use the Cost tab to recalculate that usage when you choose.
The book is a file on this computer and never leaves it.
Loading...
Tip: Ctrl/⌘+Enter to run
All runtimes on this node
How can I help?
Ask about your agents. Turn an answer into a dashboard you can keep.
Checking your assistant...
Start with a question
Ask for a decision, an explanation, or a panel for Home.
Loading...
📊 Cron Health Monitor
Click row to expand run history
Loading health data...
Memory
Markdown files your agent reads and updates. Edit them too if you want.
Summary
All files
Access log
Loading...
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...
Loading...
📊 History
Token Usage Over Time
Cost Over Time
Active Sessions
Cron Runs
Loading...
Messages / min0
Actions Taken0
Active Tools0
Tokens Used -
Tool calls follow the selected runtime. This is a limited preview. Tokens describe the node's primary session.
Connecting to recorded activity...
How your messages get answered
You
send a message
Channels
0/min
Gateway
routing
Brain
model
Tools
idle
Reply
replies back to you
📡 Active Session Lanes
Loading…
🔧 Live Tool Call Stream
0 events
Waiting for activity...
🧠 Brain · Unified Activity Stream
Activity density · last 60 min (30s buckets)
Live event stream (newest first)
Loading recent activity…
🔄 Self-Evolve
How your agent could improve.
Structured review of recent activity. Each finding has evidence and a concrete action.
📬 Notifications
Pro
One place to manage every delivery destination, used by Alerts and Approvals.
Tracing
Every session as a trace. Open one for the span tree, waterfall, and agent graph - with full inputs/outputs per span.
Loading traces…
Agent Graph
See which agents started other agents. This view uses recorded spans to connect their sessions.
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.
Loading tool catalog…
See how much of each context window your agents use. Review compactions that occur before or after the window is full. See how many tokens each compaction removes and which sessions repeatedly reach the limit.
Loading utilization…
Harness
Tools, memory, and control rules that help a model perform tasks.
What is a harness?
A model processes input and produces a response.
The harness connects the model to tools, memory, and control rules.
It can call the model repeatedly to perform a task.
Agent runtimes such as Claude Code and Cursor provide a harness.
ClawMetry shows their recorded activity.
Use the links below to inspect each part.
(Read about harness engineering)
🔁 The loop
An agent can read input, choose an action, use a tool, and check the result. This sequence can repeat until the agent finishes or reaches a limit.
Watch it live in Activity →
🔧 Tools
Tools let an agent read or edit files, run code, search the web, and call services. The available tools depend on the runtime and its permissions.
See its tools →
📚 Memory
A harness can save information from earlier sessions. It can supply this information to the model in a later session.
Read its memory →
🪟 Context
A model can process a limited amount of text at once. The harness manages that text and can remove older information.
See context usage →
📦 Sandbox
A sandbox limits the files, network connections, and other resources that agent code can access. The limits depend on its configuration.
Check its sandbox →
🚦 Guardrails
Rules define which actions an agent can take and which actions need approval. Examples include file deletion and spending limits.
Review approvals →
🤝 Teamwork
An agent can give parts of a task to other agents. It can then combine their results.
See who spawned whom →
💬 Interfaces
A runtime can provide access through a terminal, a browser, or a messaging service such as Telegram or Slack.
Follow the message flow →
Is this repo ready for an agent?
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…
Loading swimlanes...
🛡️ Security
?
Security Posture
Scanning configuration...
-
Passed
-
Warnings
-
Failed
🔑
Build your own view
Give a page, script, or coding agent permission to read selected agent data from this machine. This access cannot start, pause, or stop an agent.
What may it read?
Copy this now. It is not shown again.
🔒
Tamper-evident log
Checking the activity log for tampering...
🗄️
Data retention
Checking how long this machine keeps activity history...
days
Threat Detection & Anomaly Alerts
0
Critical
0
High
0
Medium
0
Clean Sessions
Threat timeline (newest first)
Scanning...
🗂️ Recorded findings
Recorded findings remain available after a scan. They include reports from connected security tools. Open a row to read its source session.
Loading findings...
🔒 Policy Events (PII · Injection · Credential Leak)
-
PII
-
Injection
-
Cred Leak
Scanning...
🔑 API Key Scan
scanning...
Scanning session events for leaked API keys...
📋 Signature Catalog ▶
📋 Recent activity (approvals, budgets, pauses)
No recorded activity yet. Approval decisions, budget changes, and pauses appear here.
Loading tool policy…
Loading approval audit…
Approvals ask for a decision. Review the requested action and its reason before choosing a response.
Approvals
Pending Approvals
Loading...
Decision history
Recent Decisions
No history yet.
Approval rules and delivery settings
Protection Rules
Where approvals go
Loading...
🤖 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... | |||
🔒
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.
No credit card required. Cancel anytime.
🆕 Skills
Shortcuts your agent can use. Click a skill to browse its files.
Loading...