ClawMetry for enterprise
See what AI agents cost. Control what they can do.
Bring visibility to supported coding agents and instrumented AI applications. Investigate spend, apply supported runtime controls, and give platform and security teams evidence they can review.
Want to look before a call? Open the illustrative demo: a read-only sample dashboard filled with invented data, with no sign-in. It is not customer data, and it does not show that a control ran on a real agent.
Calendar blocked by your company? Email enterprise@clawmetry.com and we will propose times.
01Understand AI spend
ClawMetry records cost per session, model and runtime on the machine where the agent runs, and attributes it by agent, skill, tool and team mapping where the runtime records enough to do so. Three numbers are kept apart, because they answer different questions.
| Number | How it is produced | Use it for |
|---|---|---|
| Estimate | Derived at ingest from the recorded token split (input, cache write, cache read, output) and a multi-provider pricing table. The session is marked estimated. | Runtimes that write tokens but no dollars. Trends, comparisons and finding the expensive runs. |
| Exact | Dollars or credits the runtime itself records, or a session where every call is priceable. Marked exact. | Runtimes that write cost on disk. |
| Contract or billed amount | Your provider contract, discounts and invoice. ClawMetry does not read invoices or contract rates, so it never presents an estimate as a bill. | Finance reconciliation, done against your provider's billing. |
Where team AI spend goesHow cost is computedAttribution by runtime, model, agent and team
02Protect delivery margins
On a fixed-fee engagement, AI usage is a delivery cost the client price does not move with. When it runs over plan, the difference comes straight out of margin, and it often surfaces on an invoice after the work is delivered.
Illustrative arithmetic, not customer data
- A fixed-fee project is priced at $100,000. Planned delivery cost is $80,000, including $4,000 of AI usage. Planned margin: $20,000, or 20%.
- AI usage lands at $10,000 instead, for example from retries on a failing build and long sessions re-reading a large repository.
- Delivery cost becomes $86,000 and margin falls to $14,000, or 14%. The client price did not change.
Seeing spend by agent, session and team mapping during delivery leaves time to change how agents are used. ClawMetry shows where the number comes from. It does not guarantee a lower one.
03Investigate avoidable consumption
- Repeated failuresDetectors flag stuck loops, repeated tool failures and runs that make no progress, with the estimated spend of the flagged stretch. Where cost is unknown the figure is zero and says so. Detectors
- Large contextsContext views show what an agent re-reads on every turn, such as whole files where a search would do. Reducing spend
- CachingCache hit rate and cache-risk views show sessions where a changed prompt prefix throws away a large cached prompt.
- Expensive runsSort sessions and models by cost to see which runtime, model and session drove it. Attribution
A recommendation is a hypothesis. A cheaper model, a trimmed context or a reordered prompt only saves money if the work still gets done. Agree a task-quality measure, compare before and after, and keep the change only when an experiment shows quality held and cost fell.
04Govern risky actions
Five different controls, with different timing and different requirements. Observation never needs the others; each step down the table acts earlier and works in fewer places.
| Control | What it does | Where it works | Limit |
|---|---|---|---|
| Observe | Records sessions, tool calls, tokens and cost from each runtime's own files or telemetry. Changes nothing. | Every runtime on the coverage list, to the depth that runtime records. | A runtime that writes no model id or cost shows those fields as unknown. Coverage |
| Alert | Budget, velocity and detector alerts, sent to channels you configure. | Any observed runtime. | An alert arrives after the spend or the behaviour. It stops nothing. Budgets and alerts |
| Pre-tool approval | Holds a matching tool call until a person approves or rejects it. | Only runtimes that expose a pre-tool hook. Today that is Claude Code's opt-in PreToolUse hook. | No hook, no gate. Where a hook fails open, the gate is advisory. Approval gates |
| Routed budget enforcement | The optional proxy warns, refuses or downgrades a model call at a daily or monthly limit, before the tokens are spent. | Only model traffic you route through the proxy. | Traffic that does not pass through the proxy is not limited. Enforcement proxy |
| Process stop | Pause, stop or kill an agent's process tree, pressed by a person or fired by a policy. | Depends on runtime, operating system and session. A CLI agent's process tree can be controlled; a conversation inside a shared editor process cannot. | Autonomous policies are off by default and start in monitor mode. Process control |
05Connect your environment
Each environment is listed with the status we can back today. There is no logo wall: a missing integration is shown as missing.
| Environment | Status | What that means |
|---|---|---|
| Supported coding agents | Available | Read from each runtime's local files or telemetry. What is recorded varies by runtime. What each runtime exposes |
| Applications instrumented with OpenTelemetry | Available | Send OTLP over HTTP to /v1/traces, /v1/metrics and /v1/logs. OpenTelemetry ingest |
| Azure virtual machines and virtual desktops | Not tested | No configuration has been validated on Azure virtual machines or virtual desktop pools. The fleet rollout guide lists what was tested on CI runners and what was not. A pilot can test your image and report what it found. |
| LLM gateways, such as LiteLLM | Pilot scope | Since clawmetry 0.12.880 ClawMetry reads a LiteLLM proxy's own OpenTelemetry telemetry, kept as a separate subtotal at the rates LiteLLM recorded, never added to agent totals. No other gateway has a published integration, and fitting either to your keys and team layout is work a pilot would scope and test. Every integration, with its limits |
| Chat assistants used in a browser tab | Not supported | General chat products used in a browser are not observed. Runtimes with a browser interface that ClawMetry does read are listed individually on the coverage page, each with its own limits. |
Every integration, with its direction, requirements, limits and status
06Review deployment and security
How each mode handles agent data. "Cloud sync off" and "offline" are different settings, and the table says which requests remain.
| Mode | Where agent content lives | What leaves, and who can read it | Access control |
|---|---|---|---|
| Local only | A local DuckDB file on the machine, not encrypted on disk. Credential-shaped secrets are fingerprinted before they are written. | No session content. A default install still sends an install ping, a failure report and a PyPI version check, which ClawMetry and PyPI can read. CLAWMETRY_OFFLINE=1 stops all three. | The dashboard binds to loopback by default. Anyone who can reach the machine's local API can read it. |
| Managed sealed snapshots | On the machine. Prompts, replies, tool inputs and outputs are sealed with AES-256-GCM before upload, using a key created on the machine that our servers do not store. | Sealed content, plus operational metadata that syncs in readable form: hostname, per-session runtime, model, status, timestamps, token and cost totals, tool names and counts, and detector one-liners. ClawMetry can read the metadata, not the sealed content. | Hosted dashboard. SAML and OIDC single sign-on is set up per organization. Roles are owner and member. |
| Direct OTLP push | Sent to ClawMetry Cloud with a hosted ingest key. | The pushed telemetry itself, readable by ClawMetry. It is not end-to-end encrypted. | Hosted dashboard, as above. |
| Customer-hosted | On your server, run from the published Docker Compose deployment. A server that declares plaintext storage stores content unencrypted, and the CLI says so at connect. | Nothing to ClawMetry by default. An optional daily licence ping is off unless you turn it on. | A shared node token plus admin credentials. No single sign-on on the customer-hosted server. |
Architecture overview (PDF, no form)Trust hubSecurityCompliance statusSub-processorsDPA statusDeployment modes
07Start a scoped pilot
One team, the runtimes that team actually uses, and acceptance criteria written down before anything is installed.
- Scope. Agree the team, runtimes, machines or images, deployment mode, and the questions the pilot must answer.
- Readiness. Install access, network approvals and the chosen deployment mode are in place. The evaluation starts only after access and technical readiness, not at signature.
- Evaluate. We propose a four-week evaluation on the agreed machines, with a review of findings along the way.
- Read out. Results against the acceptance criteria you wrote in step one.
| Deliverable | What it contains |
|---|---|
| Cost baseline | Spend by runtime, model and session over the pilot, with the estimate or exact basis of each figure. |
| Coverage report | What was observed on each agreed runtime and machine, and what was not. |
| Control-test results | The controls you selected, tested on your runtimes, with the result of each. |
| Rollout recommendation | Whether and how to extend, including any gaps found. |
Scope and pricing are by agreement. No saving is guaranteed: the baseline shows what there is to act on. Capabilities not yet validated in the product, such as a virtual desktop image or a specific LLM gateway, are tested as pilot scope and reported as found.
08Procurement questions
Single sign-on and roles?
On the hosted dashboard, SAML and OIDC single sign-on is available, set up per organization and limited to your email domain. Roles are owner and member; finer roles and SCIM provisioning are not built. The customer-hosted server has no single sign-on.
Can ClawMetry read our prompts or our customer data?
Not the content. Prompts, transcripts, and activity content are encrypted on your machine with a key our servers do not store, and reach the cloud only as ciphertext. ClawMetry can read operational metadata: hostname, session ids, runtime and model names, token and cost totals, tool names, and short detector messages. Telemetry pushed with a hosted ingest key is readable by design.
What is the retention policy?
On the machine, events are kept for the plan's window: 7 days on Free, 30 on Starter, 90 on Pro, and unlimited on Enterprise. Daily rollups are kept after events expire, and you can set a shorter window. No retention or deletion period is published yet for data held in ClawMetry Cloud. Retention
Do you offer self-hosted or on-prem?
Yes. The OSS daemon is self-hosted by default. Enterprise customers get an on-prem license that is verified on the machine without contacting ClawMetry, and run the ClawMetry server on their own infrastructure from our published Docker Compose deployment. There is no Helm chart or Kubernetes package yet.
Which security documents are available?
The trust hub, the assurance status, the sub-processor list and SECURITY.md. Report a vulnerability to security@clawmetry.com; the SECURITY.md targets are acknowledgement within 2 business days and a fix for a confirmed critical or high issue within 30 days. They are targets, not a contractual SLA.
What licenses are involved?
OSS daemon: MIT. Pro plans: commercial subscription. Enterprise: commercial annual licence with customer-hosted rights, with scope agreed with you. Plan names and self-serve prices are on /pricing.
What support do you provide?
Support hours, contacts and escalation are agreed in writing for a pilot or contract. There is no contractual SLA for uptime today.
Who is behind ClawMetry?
ClawMetry is published by Instalabs LLC, Sheridan, Wyoming, United States. Founder: Vivek Chand (vivek@clawmetry.com). Enterprise questions: enterprise@clawmetry.com. Entity, licences and contacts: company page.
What is not available yet?
No SOC 2 report, ISO 27001 or ISO 42001 certification, or independent penetration test. No data processing agreement yet. No SCIM provisioning, no roles finer than owner and member, no single sign-on on the customer-hosted server, and no contractual SLA for uptime. No validated configuration for Azure virtual desktops or a specific LLM gateway. Anything on this list can be discussed; none of it is promised by this page.
Talk to us
A 30 minute call covers your runtimes, your deployment constraints and what a scoped pilot would test.
Book a 30 min call →Or email enterprise@clawmetry.com. Security questions: security@clawmetry.com. Reply within 1 business day.