Observability and cost tracking for OpenHands
See every OpenHands conversation, command, file edit, delegated sub-agent and token as it happens. ClawMetry finds OpenHands on your machine on its own and renders sessions, transcripts, token splits and cost in one dashboard, next to every other agent you run.
What OpenHands is
OpenHands is All Hands AI's open-source platform for software engineering agents, MIT licensed and one of the most starred agent projects on GitHub. It began life as OpenDevin in 2024 and now runs the full loop: the agent plans, edits files, runs shell commands, browses the web and iterates on its own output inside a sandbox. It is model-agnostic, so it runs on whichever model key you give it.
You reach the same agent several ways. There is a terminal CLI, the Agent Canvas browser client you can host yourself, a Python SDK for embedding agents in your own code, an agent server that exposes runs over an API, and OpenHands Cloud for teams who would rather not operate sandboxes. This page is about the ones that run on hardware you control.
What ClawMetry sees on OpenHands
Once ClawMetry is running, your OpenHands activity shows up as:
- Sessions per project: one row per conversation, labelled with the directory it worked in, the model behind it, when it started, when it last moved, message count and its token totals.
- The full transcript: your instruction, the agent's reasoning, every command it ran, every file it edited and every page it opened, each with its result, in the order they happened.
- Delegation fan-out: OpenHands agents can spawn sub-agents and hand them work, in parallel or one at a time. Each child appears under the turn that spawned it, with the task it was handed, what it answered, and its own tokens and tool calls, so a delegated branch that ate half the budget is visible rather than folded into the parent.
- Token splits that include the cache: fresh input, cached input, cache writes and output, counted per model call and added up per session, so one long agentic turn is not reported as a single call.
- Cost you can trace: OpenHands runs on your own key and records what each model call cost as it makes it. ClawMetry totals that per session, per project and per day, rather than re-deriving a number of its own.
- Condensation and retries: when the history outgrows the context window and OpenHands condenses it, or a step fails and is retried, that lands on the timeline instead of vanishing.
- Approvals: what the agent asked to do before it did it, what you answered, and which runs were started in an auto approve mode where nothing was asked at all.
- Microagents and repo instructions: the knowledge and repository microagents a conversation loaded, browsable next to every other runtime's skills and rules.
- Next to everything else: OpenHands sits in the same dashboard as Claude Code, Codex, Cursor and the rest, so fleet cost per hour is one number instead of six tabs.
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.cmd2. Run the daemon
clawmetryThe daemon finds OpenHands on the machine and starts reading immediately. No config, no restart of OpenHands, and nothing to add to your agent code.
3. Open the dashboard
Open http://localhost:8900. Switch the runtime chip to OpenHands.
Start the free trial →The one thing this catches that OpenHands alone does not
Inside a conversation you can see that conversation, and OpenHands will tell you what it has spent so far in it. What no single session shows you is the shape of the week: how many conversations you left running across how many repos, which ones burned tokens re-reading the same files, how much went to sub-agents you delegated to and forgot about, how often the context was condensed because a run would not stop growing, and how the OpenHands total compares with the Claude Code and Codex runs happening beside it. ClawMetry keeps that view: spend per hour across every session and every runtime, delegated children billed separately from their parent, repeated-call detection and token-growth curves per session. Set a threshold and the approval queue pauses before the next call. Kill switch (Pro) stops it.
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 26 runtimes ClawMetry observes
OpenHands is one of 26. Free on every plan: OpenClaw, NVIDIA NemoClaw. Paid on Cloud Pro or Self-Hosted Pro: the rest.