agent-kill-switch
/agent-kill-switch
Give the human an off switch and a cost meter for the coding agents on this machine, using ClawMetry. Use when the user asks to stop or kill a runaway agent, see what an agent is doing or spending, cap agent spend, gate risky tool calls behind approval, or "set up a kill switch" for Claude Code, Codex, Cursor, Gemini CLI, OpenClaw, Goose, Aider and similar runtimes.
- Author
- vivekchand ↗
- Repository
- clawmetry ↗
- Licence
- MIT
- Last repository change
- 26.10.2026
Assessment and security
- Files in this package
- 1
- Scripts
- 0
- HTML comments
- 0
- URLs in SKILL.md
- 9
No signals detected. Repository changed on 26.10.2026, without references to retired models or APIs.
MIT, open: use and modify while retaining author attribution.
What it does and when to use it
Give the human an off switch and a cost meter for the coding agents on this machine, using ClawMetry. Use when the user asks to stop or kill a runaway agent, see what an agent is doing or spending, cap agent spend, gate risky tool calls behind approval, or "set up a kill switch" for Claude Code, Codex, Cursor, Gemini CLI, OpenClaw, Goose, Aider and similar runtimes.
Triggers from the description
Original author description
Give the human an off switch and a cost meter for the coding agents on this machine, using ClawMetry. Use when the user asks to stop or kill a runaway agent, see what an agent is doing or spending, cap agent spend, gate risky tool calls behind approval, or "set up a kill switch" for Claude Code, Codex, Cursor, Gemini CLI, OpenClaw, Goose, Aider and similar runtimes.
How it works
- 1. Install and start
- 2. See what is running and what it costs
- 3. Stop one session (the off switch)
- 4. Hold risky tool calls until a human approves (Claude Code)
- 5. Cap spend for agents that call the API with a key
- Honest limits
- Remove
SKILL.md preview
--- name: agent-kill-switch description: Give the human an off switch and a cost meter for the coding agents on this machine, using ClawMetry. Use when the user asks to stop or kill a runaway agent, see what an agent is doing or spending, cap agent spend, gate risky tool calls behind approval, or "set up a kill switch" for Claude Code, Codex, Cursor, Gemini CLI, OpenClaw, Goose, Aider and similar runtimes. license: MIT metadata: author: vivekchand homepage: https://clawmetry.com source: https://github.com/vivekchand/clawmetry --- # Agent kill switch An agent that loops, re-reads the same files, or shells out to something it should not is expensive by the minute. This skill wires up [ClawMetry](https://github.com/vivekchand/clawmetry), a local, read-only dashboard that reads the session logs the runtimes already write on disk, and uses it for three things: see what is running, stop one session, and hold risky tool calls until a human says yes. Nothing here modifies the agent. Observation is read-only. Every control below is opt-in and stated as what it does, so tell the user exactly which ones you enabled. ## 1. Install and start ```bash pip install clawmetry && clawmetry ``` The dashboard is at `http://localhost:8900`. It auto-detects the runtimes on the machine (Claude Code, Codex, Cursor, Gemini CLI, OpenClaw, Goose, Aider, Cline, OpenCode and others); no config. Confirm it is up: ```bash curl -s http://localhost:8900/api/overview | head -c 400 ``` If the port is taken: `clawmetry --port 8901` and use that port below. ## 2. See what is running and what it costs ```bash curl -s http://localhost:8900/api/sessions # live and recent sessions, ids look like claude_code:<uuid> curl -s http://localhost:8900/api/usage # tokens and cache-aware cost per model and per session ``` Tell the user in plain words: which sessions are active, which one is spending the most, and whether any is showing loop signs. The dashboard's Sessions tab shows the full transcript and every tool call for a session. ## 3. Stop one session (the off switch) ```bash curl -s -X POST http://localhost:8900/api/sessions/<session_id>/stop ``` - `200` with `"ok": true` and a `pid`: the session's process received SIGTERM. - `409` with `"ok": false`: ClawMetry could not identify a process for that session, for example a Cursor editor conversation, which shares the IDE process. Say so instead of claiming it was stopped. Never call this on your own initiative. Only when the user asks to stop a session, and confirm which session id first. ## 4. Hold risky tool calls until a human approves (Claude Code) ```bash clawmetry hooks install # adds a PreToolUse hook to ~/.claude/settings.json, idempotent clawmetry hooks status ``` Then one policy gates everything ClawMetry scores `high` or `critical` (recursive deletes, force pushes, sudo, credential files, reverse shells, cloud metadata endpoints): ```yaml # ~/.clawmetry/policies.yml - name: 'Require approval for high-risk actions' min_risk: 'high' action: 'require_approval' timeout: 604800 on_timeout: 'deny' ``` A deny blocks that one call; the agent keeps its session. Approvals can be answered from the Approvals tab or, with cloud sync, from a phone push. Undo with `clawmetry hooks uninstall` (removes only ClawMetry's entries). ## 5. Cap spend for agents that call the API with a key ```bash clawmetry proxy start --daily-budget 10 export ANTHROPIC_BASE_URL=http://localhost:4100/v1 # or OPENAI_BASE_URL for OpenAI-backed agents ``` When the cap is hit the proxy returns `BUDGET_EXCEEDED` instead of forwarding the request. It also detects repeated near-identical requests (loops). This only covers agents whose traffic you route through it; say that. ## Honest limits - OpenClaw and NemoClaw are free in the open-source app. The other runtimes, and the approvals feature, need ClawMetry Cloud or a self-hosted Pro license after the trial. Tell the user before enabling those. - The stop in step 3 is SIGTERM to a process ClawMetry can identify. It is not a guarantee for every runtime; the `409` is the honest answer. - Data stays on the machine unless the user runs `clawmetry connect`, which enables end-to-end encrypted cloud sync. ## Remove ```bash clawmetry hooks uninstall clawmetry proxy stop pip uninstall clawmetry ```
Versions
The database has no other skills named agent-kill-switch.
Identical content appears in 1 other GitHub folders. We show one copy with the highest score.
Database snapshot: 13.11.2026 · package version 62ac9057d3a9 (SHA-256 digest of the entire package)
Downloads in the last 30 days: 0 · ranking
Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.
Similar skills
- 100agent-oversight
Adds runtime oversight to an AI agent so it cannot do more than its AI_GOVERNANCE.md allows - default-deny tool access, amount and call limits, human approval for risky actions, spend limits, a kill switch, prompt-injection trust boundaries and a decision log - using the… - 95clawmetry-selfcheck
Read your own agent telemetry from ClawMetry (waste, progress, cost) and act on it before finishing a task. Use when ClawMetry is installed on this machine and you want to check whether you are re-reading files, spinning in loops, or burning budget. - 100bootstrap-system-md
Explore a codebase and generate its first SYSTEM.md, a single tool-agnostic living reference any coding agent (Claude Code, Codex, Cursor, Aider, Gemini CLI) reads before working on the repo, or upgrade an existing SYSTEM.md to the current contract. Use when the user asks to… - 90handoff
Keep a handoff file so another coding agent (Claude Code, Codex, Kimi Code, Gemini CLI, Cursor) or a later session can continue this task without re-explaining. Use when the user asks to start, update or resume a handoff, says they will switch agents, when… - 100end-to-end-kill-switch-audit
Use when verifying that a kill switch or send cap works link by link across a multi-process chain, instead of trusting an aggregate counter that did not move. - 10014-ai-agent-architecture-spec
Use when an approved tool-using AI agent needs a runtime loop, state machine, memory tiers, planner, dispatcher, durability, kill switch and tenant isolation; use AI architecture for non-agent AI services and multi-agent coordination only when several agents share a task.