在你的電腦上執行 · 只在你允許時行動
AI 代理的
控制中樞
- 看見每一個代理。
- 理解每一個動作。
- 阻止不該發生的事。
你的電腦有 System Monitor 來監看程序，這一個則是為 AI 代理 而生。代理一啟動就自動偵測，每個 token 都對應到花掉它的任務，每次失控迴圈都在燒光預算前就被標記出來。
免費試用，無需信用卡。 macOS (.dmg) Windows (.exe) Linux (.AppImage)
或透過終端機安裝
$ pip install clawmetry && clawmetry
7 天全執行環境試用，免綁信用卡。 之後每個節點每月 $9 或 $19。
“claude_code: 84 tool calls, no file changes, not advancing.” Stop or Pause this agent from the ClawMetry dashboard or device.
Verbatim record, not a mockup. A real detection from the laptop ClawMetry was built on. The steps, the counts and the message are what the daemon wrote to disk.
ClawMetry 存在的理由
每個 AI 代理都需要一個終止開關。
當自主系統失控時，誰握有控制權？這是 ClawMetry 背後的故事，從兩部童年電影，到一個獨立的控制平面。
閱讀創辦人的故事 →
❯ Supported agent runtimes
Don't see the agent harness you use?
Just share it and we'll add support in ~2 days. Leave your email and we'll ping you the moment it's live, plus 2 months of Pro, on us.
We only use your email to notify you and issue the Pro credit. No lists, no spam.
實測結果，單一安裝，為期四個月
它標記的是停滯的代理，
不是花錢的代理。
一段漫長又昂貴的工作階段，往往正是你最出色的成果，所以一味盯著大數字的工具只會誤殺。ClawMetry 看的是代理究竟有沒有產出。同一台筆電，四個月，兩種情況並列對照。
運作正常 · 從未被標記
$171.58
The most expensive session of the four months. It ran long and cost real money, and it was writing code the whole time. ClawMetry left it alone, which is the correct answer.
已標記 · 停滯中
$175.10
Every one caught for the same reason: tool calls with no file changes, the same failure repeating, or a loop that stopped advancing. Most of them cost less than the session on the left.
Detectors warn. They do not act. All seven of those sessions ran to completion. The flag is a notification, not a trigger. Nothing is stopped unless a person stops it, or unless you set a spend cap yourself and an agent crosses it.
165 sessions and $826.47 of agent spend on one developer’s laptop, April to August 2026. One machine, not a customer aggregate. Yours will look different, which is the reason to read your own.
實際運作一覽
一個畫面，掌握所有代理的即時動態。
缺口就寫在表格裡
各執行環境能做到什麼。
這個領域的每家廠商都在談防護機制。這是完整的對照表，連空白格也照列。
| 執行環境 | 觀測 | 執行前攔截 | 執行中終止 | 美元成本 |
|---|---|---|---|---|
| OpenClaw, NVIDIA NemoClaw, Goosefree forever | ✓ | via proxy | ✓ | ✓ |
| Claude Codethe only pre-execution tool gate | ✓ | ✓ | ✓ | ✓ |
| Codex, OpenCode, Aider | ✓ | via proxy | ✓ | ✓ |
| Qwen, Pi, Grok Build, DeepSeek Harnessper-terminal CLI processes, resolved by argv and cwd | ✓ | via proxy | ✓ | ✓ |
| GitHub Copilotbilled in vendor credits, not USD | ✓ | – | ✓ | – |
| Kimi CLIno model id on disk | ✓ | – | ✓ | – |
| Hermes, Deep Agents, n8n, Antigravity, QM, Exo |
✓ | via proxy | – | ✓ |
| NanoClaw, PicoClawno model id on disk | ✓ | – | – | – |
| Cursorone IDE process holds every session | ✓ | – | – | – |
Eight of the 30 have adapters but are not yet audited into this table: Gemini CLI, Lovable, Cline, OpenHands, OpenWorker, Devin, Grok Bot and Replit Agent. They are observed; the other three columns are unverified or unavailable, so they are absent rather than guessed.
執行前攔截
Claude Code gets a real pre-execution gate: a PreToolUse hook parks the call before the tool fires. Everything routed through the ClawMetry proxy gets spend refused with a 429 before tokens are spent.
執行中終止
A gateway kill for OpenClaw, and SIGSTOP then SIGKILL against the process tree for the CLI runtimes, with a guard so a recycled pid is never the one we hit. macOS and Linux. Windows has no SIGSTOP, so pause and resume do not exist there.
事後稽核
Every event carries a SHA-256 chain over its immutable fields, per node, with a verifier command. Approval decisions land in an append-only log. Works on every runtime.
在你把常駐程式裝上筆電之前
它對電腦做了什麼，
又有什麼會離開電腦。
以一套參考安裝實測：165 個工作階段、19,977 筆事件。並非估算值。
資源占用
clawmetry uninstallThe query server binds to loopback only, with a one-time 32-hex token and a discovery file at mode 0600. It is why the dashboard can read the store without holding the writer lock.
確切外傳的內容
Never. These stay in the local file.
The key is generated on the node and never leaves it. Decryption happens in your browser. Compelled by a subpoena, we hand over a blob we cannot open, plus the node id, the version and some counters.
本機檔案為純文字
Everything is sealed in transit and at rest in our cloud, and it is not encrypted on your own disk. The DuckDB file carries default user permissions. Anything running as your user can read it.
遮蔽機制失效時會放行
Secrets are fingerprinted at ingest by default, before disk. It is pattern-based, and on an internal error it keeps the original value rather than dropping data. It is tuned for credential shapes, not customer PII.
已連線的節點可被遠端停止
If you turn cloud on, Stop, Pause and Resume are reachable from the dashboard and the desk device for any node on the account. That is the feature. Run local-only and no such channel exists.
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.
三種執行位置
Cloud 是選配。
一向如此。
常駐程式會把運算結果寫進你自己磁碟上的 DuckDB 檔案。之後儀表板要放在哪裡由你決定，而且無論選哪一種，價格都相同。
01 · Local only
沒有任何資料外傳
Included in the MIT core
The dashboard reads the local file through a loopback query server. There is no account, no sync and no destination to point at.
- Network destinations: none
- Inbound ports opened: 0
- Free forever for OpenClaw, NVIDIA NemoClaw and Goose
02 · Our cloud
外傳前先行加密封裝
$19 per node, per month
The snapshot is encrypted on the node with a key generated there and never sent. Your browser decrypts it. Subpoena us and we hand over a blob we cannot open.
- AES-256-GCM, key held on the node
- Decryption happens client side
- Stop, Pause and Resume reach any node on the account
03 · Self-hosted
你的基礎架構，單一租戶
$190 per node, per year
One container serves the UI and the ingest API on your own network, with node tokens and admin auth. Licences verify offline against an embedded Ed25519 key you can fingerprint and compare.
- No calls to clawmetry.com beyond an off-by-default version ping
- Same feature set, same runtimes, same rate
- Two months free against the monthly price
價格不會因部署方式而改變。自架部署每個節點每月 $19，或每個節點每年 $190，與 Cloud 方案完全一致，因為你付的是軟體的錢，不是我們伺服器的錢。改變的只是誰握有資料、誰來執行容器。直接購買自架部署授權金鑰，不必先跟任何人談。
單一費率，單一計量方式
在跟任何人接洽之前，
先算出你的數字。
7 days of Pro, every runtime, no card
$9per node, per month
Starter, billed monthly. Annual is $90 per node, two months free. Starter is the observability tier: fewer features than Pro, not fewer runtimes. Free forever for OpenClaw, NVIDIA NemoClaw and Goose.
Start the trial- Every runtime, including Claude Code, Codex and Cursor.
- Cloud or fully self-hosted, at the same rate. Unlimited chat channels.
- Cost and token split, budget limits, anomaly detection, approval queue.
- Cancel any time in Stripe. No contract, no seat count.
Not in Starter: the pre-execution tool gate, waste flags, per-run compare, error triage, Self-Evolve findings, the eval suite and OTel export. Those are Pro.
$19per node, per month
Pro, billed monthly. Annual is $190 per node, two months free. Pro is the governance tier: it adds gates, evals and waste analysis on top of Starter. Free forever for OpenClaw, NVIDIA NemoClaw and Goose.
Start the trial- Everything in Starter, and the same 30 runtimes.
- Tool policy with a pre-execution gate, and an eval suite that scores runs.
- Waste flags, per-run compare, error triage, cost optimizer, OTel export.
- Cancel any time in Stripe. No contract, no seat count.
The 7-day trial is Pro, whichever plan you land on afterwards. No card, and nothing is charged if you stop.
Already decided? Buy now and install after. Your key is on the next screen. Compare Starter and Pro.
你的花費是多少
計算會執行常駐程式的機器數量，而不是會查看儀表板的人數。
$551 $6,612 per year · 29 nodes
A node is a machine running the daemon and reporting a heartbeat. The unit of attribution is a machine, so spend maps to a laptop, not to a team. The same number applies self-hosted, so you can price a deployment on your own infrastructure off this figure without asking us. Ephemeral CI runners are not yet specified; the fleet brief carries the rule in writing before you sign.
為承擔風險的人而設
在整間公司導入。
上一季，你的工程師裝了 Claude Code、Cursor、Codex 和 Copilot。沒有人走過採購流程，而它們全都握有你的程式庫和你的金鑰。
無需在代理端安裝任何東西
Each runtime is read from the session store it already writes. No SDK, no wrapper, no second agent on the laptop, nothing a developer has to opt into or can forget.
A control layer that governs only its own SDK is blind to everything a developer runs outside it.自架部署就是 SSO 的解答
SSO ships: SAML and OIDC through WorkOS, with the pinned domain gating provisioning. Roles are owner and member, and nothing finer. If your rollout needs more than that, it is not built yet. Single-tenant self-hosted runs air-gapped: one container serves the UI and the ingest API, with node tokens and admin auth, and licences verify offline against an embedded Ed25519 key you can fingerprint and compare.
A self-hosted deployment makes no calls to clawmetry.com beyond an off-by-default version ping.契合你現有的網路架構
SIEM and syslog export in CEF or JSON over TCP, UDP or TLS, emitted daemon-side because the cloud never holds plaintext. Corporate TLS interception is supported, with a doctor command that walks DNS, TCP, proxy and TLS in order.
In: your collector feeds us over OTLP. Out: full-event SIEM export. Not yet: correlated spans carrying your trace ids.機群部署目前仍需手動進行
There is no Helm chart, no Terraform provider and no Intune or Jamf package. If a daemon on developer laptops is a non-starter, run the enforcement proxy centrally and govern spend without touching an endpoint.
Auto-update ships on by default. CLAWMETRY_AUTO_UPDATE=0 turns it off. No staged rollout, no version pinning.每次都會被問到的問題。
可以在執行前攔截，還是只能事後告訴我已經執行了？
Both, and the split is in the table above. Claude Code gets a genuine pre-execution tool gate, and anything routed through the proxy gets spend refused with a 429 before tokens are spent. For the other runtimes it is detection plus process control: we can see it and end it, we cannot veto the call it already made.
Is ClawMetry only for OpenClaw?
No. It launched in February 2026 as an OpenClaw-only dashboard and many directory listings still describe it that way, but it now reads 30 runtimes: OpenClaw, Claude Code, Cursor, GitHub Copilot, Codex, Gemini CLI, Lovable, Cline, OpenHands, OpenWorker, OpenCode, NVIDIA NemoClaw, Aider, Goose, n8n, Qwen, Devin, Antigravity, Grok Build, Grok Bot, Hermes, Pi, Deep Agents, Kimi CLI, DeepSeek Harness, Exo, NanoClaw, PicoClaw, QM and Replit Agent. It auto-detects whichever of those are on the machine, so a mixed fleet lands in one dashboard. OpenClaw, NVIDIA NemoClaw and Goose are free forever in the open source package; the rest are on the paid tier.
它會自行停掉我的代理嗎？
No. Detectors write a warning and stop there. On the reference install every flagged session ran to completion. A run ends only when a person ends it, or when you have set a spend cap yourself and an agent crosses it.
它會改變我的代理運作方式嗎？
Read-only by default. Two things you opt into change that: the proxy, a separate process you point one env var at, and approval gates, where ClawMetry adds a gate to the runtime’s own permission system and removes only what it added.
試用期結束後會怎樣？
The paid runtimes stop being read and the free three keep working. Entitlements are a commercial boundary, not a technical one. We do not hard-block a running fleet on a licence check.
團隊是誰？如果你們消失了怎麼辦？
One person, plus contributors. The core is MIT and runs on your machine, so if we vanish your install keeps working and your data stays in a DuckDB file you own. The paid layer is a separate closed package. Ask for the fleet brief and it names the legal entity, the largest deployment we run today, and the reference count, with no rounding.
這能通過資安審查嗎？
For a pilot on non-production machines, usually: MIT source you can audit, local-first storage, single-tenant self-hosted and offline licence verification are the arguments that work. For a regulated production rollout, read the gap list first. SSO is there; RBAC beyond owner and member, SOC 2, CAIQ and a third-party pen test are not, and that is a two-to-four quarter conversation.