How do I monitor Exo agent conversations?

Install ClawMetry (pip install clawmetry). The daemon auto-detects Exo workspaces and renders every conversation, tool call, model, and token in one dashboard at http://localhost:8900. Each Exo conversation is one session; forks and rewinds show as child lanes under the parent.

Exo forks and rewinds conversations. Does that double-count cost?

No. A forked Exo conversation carries its parent's history forward, and ClawMetry bills that history once, to the parent. The fork shows only what it spent after the fork point, so fleet cost adds up instead of inflating with every branch.

Does ClawMetry use Exo's own cost figures?

Yes. Exo records the model, token counts, and a USD cost on each model call, and ClawMetry uses that native figure when it is present. When Exo could not price a call, ClawMetry prices the tokens itself and marks the session as estimated rather than exact.

Is ClawMetry safe to point at a running Exo install?

Yes. ClawMetry only observes. It never writes to Exo, never talks to a running exo serve, and a conversation it cannot read is skipped rather than mis-reported.

How much does ClawMetry cost for Exo?

Free forever on your own machine. Cloud starts at $9/node/mo (Starter) or $19/node/mo (Pro).

Exo × ClawMetry

Exo AI observability and cost tracking

See every Exo agent conversation, fork, tool call, and token as it happens. ClawMetry detects Exo (exoharness) workspaces automatically and renders conversations, tokens, native per-call cost, and fork lineage in one dashboard.

Supported runtime Zero config, auto-detected Read-only Free trial Official Exo site ↗

📦 809k+ installs 🌍 129+ countries ⭐ 330 on GitHub 🏆 #5 Product Hunt

See ClawMetry with Exo

What ClawMetry sees on Exo

Once ClawMetry is running, your Exo activity shows up as:

Conversations as sessions: one row per Exo conversation, with the agent it belongs to, started at, last activity, model, message count, and per-session token totals.
Live activity: every tool call and result as it lands, with arguments, plus the model's reasoning summaries. A tool call shows once, even though Exo journals it in two places.
Per-turn tokens: prompt, completion, cached, cache-creation, and reasoning token counts on every model call, plus time-to-first-token and call duration.
Native cost, honestly labelled: Exo prices each model call itself, and ClawMetry shows that exact figure. When Exo could not price a call, ClawMetry prices the tokens and marks the session estimated instead of pretending.
Fork lineage without double billing: a forked conversation nests under its parent as a child lane, and the history it inherited is billed once, to the parent.
Any executor: whether the conversation ran on Exo's default executor, Codex, Claude Code, or Cursor, it lands in the same view with the same numbers.

Read-only by design: ClawMetry never writes to an Exo workspace and never talks to a running exo serve. It reads what Exo already recorded, after the fact.

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.cmd

2. Run the daemon

clawmetry

The daemon auto-detects Exo workspaces on the machine and starts reading immediately. No config, no runtime restart.

3. Open the dashboard

Open http://localhost:8900. Switch the runtime chip to Exo.

Start the free trial →

The one thing this catches that Exo alone does not

Exo is built to fork, rewind, and let an agent rewrite itself mid-run, and exo repl shows you one conversation at a time. That is exactly when spend stops adding up in your head: a branch inherits its parent's whole history, a self-edited agent starts calling tools in a loop, and three executors share one workspace. ClawMetry watches the whole fleet: cost per hour across every Exo conversation, fork trees with the inherited history billed once, message-repeat detection, and per-session token-growth curves. 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.

Peter SteinbergerOpenAI

Creator of OpenClaw and founder of PSPDFKit, now at OpenAI.

Matt Van HornJune · Weber

Co-founder of June, the smart oven acquired by Weber. ClawMetry contributor.

Alex KrentselGoogle

Systems research engineer at Google and PhD student at UC Berkeley.

Manoel Aranda NetoPostHog

Software engineer at PostHog, previously mobile SDKs at Sentry.

Kody KendallLlamaPress AI

Co-founder of LlamaPress AI, an open-source AI web-app builder.

All 31 runtimes ClawMetry observes

Exo is one of them; the other 30 are listed below. Free on every plan: OpenClaw, NVIDIA NemoClaw, Goose. Paid on Cloud Pro or Self-Hosted Pro: the rest.

Every runtime on one page, with what each one can and cannot show →