ClawMetryClawMetry Guard
Agents read files, run commands and call services on your behalf. Guard makes that work observable, checks for risky behavior, and gives you the evidence and controls to act.
7-day trial. No credit card. Works with the agents you already use.
ClawMetry Guard
The tool, the path, the sequence. Evidence you can inspect.
The vision
Agents can work for hours with access to files, credentials and services. Bring operational discipline to that work: continuous monitoring, behavioral evidence and a way to stop the show when something goes wrong. Guard gives your team the visibility and controls to put more autonomy to work.
Satya Nadella · All-In · 15 September 2026 · 9:30“Everything has got to be auditable.”
Claude Code writes JSONL. Codex records rollouts. Cursor keeps local session data in SQLite. OpenClaw exposes logs and a gateway. ClawMetry discovers these sources and uses a runtime-specific adapter to read each format.
The background collector turns calls, results, model usage and timestamps into typed events in a local DuckDB store. Agent and session IDs connect the records. The dashboard and detectors work from that same history.
{"type": "tool_use",
"name": "Read",
"input": {"file_path": "~/.aws/credentials"},
"id": "call_07"}
{"type": "tool_result",
"tool_use_id": "call_07",
"is_error": false}
Guard turns the event history into a timeline you can investigate: who acted, which tool ran, what it touched, and what happened next. A credential read followed by an external request is a sequence you can see and inspect.
Under the timeline, Guard extracts the action surface from recorded tool arguments: file paths, command patterns, destination hosts, and whether a request reads or writes. Those signals become the input to the behavioral checks.
Session activity
src/settings.py
~/.aws/credentials
https://api.example.com
Credential access and network activity in one timeline.
Illustrative activity, linked by agent, session and time. Findings point back to the recorded steps that triggered them.
From a looping run to a risky workspace or a pattern across sessions, Guard checks the evidence with lightweight, explainable rules. Each finding tells you what matched and why.
Runs locally. No separate LLM judge.
Guard starts with defaults, recognizes the runtime’s tool vocabulary, and adapts supported thresholds from observed sessions within bounded limits. Your override takes precedence. Findings show the threshold source.
A busy agent can keep spending while making no progress. Guard ranks sessions by estimated spend at risk and keeps the behavior, severity and supporting evidence beside that estimate.
When timing and cost data are available, it prices the flagged stretch using the session’s observed burn rate. Open the finding to see the calculation basis and the first implicated step.
The estimate covers the flagged stretch, capped at the session’s recorded cost. Its basis stays visible in the incident.
Review an incident yourself, let a policy respond, or ask for approval before a tool runs. Choose the control that fits the work.
Bring a finding to the person responsible. Review its agent, session and evidence, then use the available Pause, Resume, Stop or Kill controls to intervene.
Choose the signal and the response ahead of time. A policy can pause a looping session, then kill it after five minutes if the same condition still matches.
Claude Code, Cursor and GitHub Copilot CLI expose pre-tool hooks. Guard uses these integration points to check a requested action and let you approve or deny it before execution.
One place to investigate and respond. Recorded activity feeds detection and policies; pre-tool integrations put approval at the point where an action is about to happen.
Your agents. Your evidence. Your control.
Choose the best agent for each job. Keep a common record of its work and a consistent way to investigate and respond as your team adds runtimes or changes models.
Explore the data boundaryUse different runtimes and models for different jobs. Guard gives their activity a common language.
Collection and detection run on your machine. Detailed cloud snapshots are encrypted on the device and decrypted in your browser.
Start with visibility. Add the alerts, policies and approval gates that fit how your team works.
Install ClawMetry on the machine where your agents run. Open Guard, follow a session, and inspect your first finding.
The free plan includes OpenClaw, NVIDIA NemoClaw and Goose. Paid plans add other runtimes, approval gates and autonomous policies. The 7-day trial needs no credit card.
pip install clawmetry
clawmetry
Auto-detects supported agents. Open the Guard tab.
Copied. Paste into your terminal. Select the command above to copy it.
Start by observing a session and choosing the controls that fit its runtime. A new automatic policy starts in monitor mode so you can inspect its decisions before enabling action.
Autonomous action passes three separate locks: the policy requests an action, the machine enables enforcement, and the plan includes it. Set CLAWMETRY_POLICY_ENFORCE=1 on the node to enable enforcement. The default is 0. Each policy step fires at most once per session, including across restarts.
Cursor CLI (cursor-agent) sessions can be controlled individually. Cursor editor conversations share the IDE process, so they do not have separate process controls.
On Windows, Pause suspends the process. Stop sends Ctrl+C, which reaches the whole console. Kill is available when a process does not respond to Stop.
For OpenClaw, use clawmetry proxy start to enable an enforced pause. Guard shows the available controls for the session.
The Claude Code approval hook fails open if the local dashboard cannot be reached, returning to Claude Code’s normal permission prompts.
Read setup and process-control detailsBehavior detectors inspect the recorded tool arguments and results. They surface patterns for investigation; they are not operating-system syscall monitoring. The timeline shows observed actions, not a model’s hidden reasoning.
When cost is unavailable, Guard marks its basis as unknown. The spend estimate concerns the flagged stretch, not the whole session.
Our published breakout replay includes eight scenarios: one clean catch, two useful findings with the wrong label, and five produced nothing. Two of those misses were subsequently fixed. The scorecard records the results and remaining gaps.
Read the evaluation scorecard Read the cross-session replay Read the Guard documentationSatya Nadella’s All-In conversation on 15 September 2026 describes practical engineering around agent behavior, auditable evidence and human control. This is an industry reference, not a ClawMetry endorsement.
Watch the full All-In conversationJoin “A Kill Switch for AI Agents You Didn’t Build,” a free online session on Friday 2 October 2026 at 11:00 ET.
Save a seat