Observability and governance for QM agents

See every qm turn, scope, and token as it happens. ClawMetry detects qm automatically and renders sessions, tokens, cost, and the org-scope layer in one dashboard. No qm code change.

What ClawMetry sees on qm

qm (github.com/yc-software/qm, qm.ycombinator.com) is Y Combinator's open-source multiplayer agent harness: one Slack + web surface per org, per-scope sandboxes, and a central turn loop that delegates to Pi, OpenCode, Codex, and Claude Code. Once ClawMetry is running, your qm activity shows up as:

• Sessions per scope, surface (slack, web, cli), thread reference, channel name, turn count, and last activity.
• Every entry: user messages, assistant turns, tool calls, tool results, and handoffs to child harnesses.
• Token cost per session: normalised across Anthropic and OpenAI usage shapes, split into input, output, cache-read, and cache-write tokens.
• Per-turn timing and rollups: TTFT, total ms, model calls, tool calls.
• The org-scope layer: scopeId, surface, threadRef, channelName. That Pi, OpenCode, Codex, and Claude Code cannot see individually.

Enable it in about two minutes

1. Install ClawMetry

Pick the line that matches your OS. Windows users: paste the install.cmd line, not the shell one.

# macOS or Linux
curl -fsSL https://clawmetry.com/install.sh | bash

# Windows (cmd.exe or PowerShell)
curl -fsSL https://clawmetry.com/install.cmd -o install.cmd && install.cmd

Already have Python? pip works too; onboard is the interactive first-run wizard.

# Any OS with Python 3.9+
pip install clawmetry && clawmetry onboard

Or skip the terminal entirely and grab the desktop app: macOS (.dmg) · Windows (.exe) · Linux (.AppImage). It installs and updates ClawMetry for you.

2. Connect the daemon

clawmetry connect

3. Point at qm's database

Set CLAWMETRY_QM_DATABASE_URL; ClawMetry falls back to DATABASE_URL if it is unset. Point it at a read replica if you prefer.

export CLAWMETRY_QM_DATABASE_URL="$DATABASE_URL"

Open http://localhost:8900. Switch the runtime chip to QM. Every qm session, scope, and cost is there.

Start the free trial →

Safety: read-only by design

ClawMetry observes qm read-only. It is built so it cannot modify qm's data, and so a slow query on ClawMetry's side cannot slow qm down.

If your policy is stricter, point CLAWMETRY_QM_DATABASE_URL at a read replica. ClawMetry never needs write access to qm.

What each tab shows for qm

Tab	What it shows	Scope
Sessions	Every qm session, sorted by last activity, with scope, surface, and token totals	All scopes, or filter by one
Brain	Live entries as they land: user messages, tool calls, handoffs to child harnesses	Whole node, or one session
Cost	Per-session, per-scope, per-hour token spend; cache-read vs uncached-input split	Runtime = qm, or all
Sub-agents	Which child harness (Pi, OpenCode, Codex, Claude Code) served each qm turn	Per session

How this compares to running qm alone

qm's own admin surface shows current runs and audit logs. What it does not show is real-time cost per hour across scopes, cache hit rate per session, or a kill-switch when a scope's turn is token-maxing. Those are the things a team running qm on a real budget needs on day one; they are what ClawMetry adds. See the cost-per-hour visibility page and the token-maxing page for how these work end to end.

Starred by the people building AI

Engineers and founders from OpenAI, Google, PostHog and more have starred ClawMetry on GitHub. One of them helped build it.

All 20 runtimes ClawMetry observes

qm is one of 20. Free on every plan: OpenClaw, NVIDIA NemoClaw. Paid on Cloud Pro or Self-Hosted Pro: the rest.

Individual per-runtime pages are rolling out over the next few weeks. Ping vivek@clawmetry.com if there's one you want prioritised.