---
title: Runtime support
description: All {{RUNTIME_COUNT}} agent runtimes ClawMetry observes — OpenClaw, Claude Code, Codex, Cursor, Copilot, Gemini CLI, Goose, Cline, OpenHands and more — and what each exposes.
keywords: supported AI agent runtimes, Claude Code observability, Cursor telemetry, Codex monitoring, Goose agent, OpenHands, Cline, agent runtime support
eyebrow: Runtimes
---

# Runtime support

ClawMetry reads **{{RUNTIME_COUNT}} agent runtimes**. {{FREE_RUNTIME_COUNT}} are
free in the open-source package; {{PAID_RUNTIME_COUNT}} unlock on a paid tier.
Every one of them is read from its own on-disk (or, for two, remote) store —
there is no SDK to import and nothing to instrument.

{{RUNTIME_CARDS}}

## The full list

{{RUNTIME_TABLE}}

Two runtime ids exist in the entitlement list but share another runtime's read
path: `grok_bot` is read through the Grok adapter. Adapters also ship for
Lovable and Replit.

## What "supported" means here

A runtime being on this list means an adapter exists that reads its real store
and maps it into ClawMetry's unified event schema. It does **not** mean every
runtime exposes the same things — they genuinely do not, and the difference
matters:

- **Cline** records dollars, a per-message token split, process ids, exit codes,
  sub-agent lineage, real cron schedules with per-run cost, and even the tool
  calls it rejected. It is the richest store in the matrix.
- **PicoClaw**'s message struct has no token field at all. ClawMetry reports
  zero tokens and does not advertise cost for it, because the data is not there.
- **Cursor**'s backend is proprietary, so token counts come only from a hosted
  VM's loopback usage log — which is why a Cursor row can show real cost on a
  managed node and none on a laptop.
- **Kimi** persists a complete token breakdown but never the model id, so cost
  is marked `estimated` with the model resolved from config.

[What each runtime exposes](/docs/runtimes/coverage/) has the comparison table.
Each runtime's own page has the exact paths, environment overrides and traps.

## Adding a runtime that is not here

Three options, in increasing order of effort:

1. **OpenTelemetry.** If the runtime can export OTLP, point it at ClawMetry's
   receiver. This is the supported path for n8n on Postgres, hosted CI agents
   and anything without a local store.
   → [OpenTelemetry ingest](/docs/runtimes/opentelemetry/)
2. **The ingest API.** Push runs and events over HTTP from code you control.
   → [Custom runtime ingest](/docs/runtimes/custom-ingest/)
3. **The interceptor.** Two lines in a Python process capture its LLM calls,
   cost, tokens, latency and errors, attributed to a named source.
   → [Track any SDK agent](/docs/runtimes/sdk-agents/)

If you want a first-class adapter for a runtime that writes to disk,
[open an issue](https://github.com/vivekchand/clawmetry/issues/new) with the
store layout — that is the part that takes the work.

## Runtime ids

The `id` column is what appears everywhere machine-readable: the entitlement
API, `clawmetry runtimes --json`, the `runtime` filter on query shapes, session
id prefixes, and per-runtime environment overrides like
`CLAWMETRY_NOPROG_TOOLS__CODEX`. Common alternate spellings are accepted as
aliases — `claude-code` and `claudecode` both resolve to `claude_code`.