Experiment tracking for Claude Code, Codex and Cursor
Your coding agent can train, evaluate and sweep, but it forgets what it tried when the session ends. Probe records the runs, metrics and decisions behind every result, so the next agent and the next teammate start from what was already learned.
Tell your coding agent:
Install Probe on this machine and its coding agents. Read https://research.prbe.ai/install.md and follow it.
Your agent can sign you up and install Probe itself. You sign in once in the browser.
What changes for you
- You'll always know how any result was made.
- Your agent won't repeat what your team already tried.
- You can pick up any experiment where it left off, or hand it off.
The problems it fixes
- Experiment messiness. Runs, logs and notes scatter across terminals, notebooks and machines.
- Collecting results. Gathering what happened takes longer than running it.
- Repeated failures. The lesson from the last attempt never reaches the next session.
- Intent to results. You see a number but not what the agent meant to test, so you can't tell why it failed.
How it works
- The work is recorded as it happens. By default the Probe daemon, a second agent beside your coding agent, records runs, notes and summaries; or your agent records its own work. Either way, the
/instrument-codeskill wiresprobe.init()andprobe.log()into a training script, an eval or a sweep, including trainers that already have their own tracker. See The Probe daemon. - Writes never block training. The SDK queues what you log and sends it in the background. See the SDK quickstart.
- Agents read before they start. Your agent asks the team's record with
probe ask, or, when it records its own work, through a read-only MCP server with six tools: browse, search, entity, metrics, query_sql and find_papers. See MCP server. - You choose per conversation.
/probe onrecords and reads,/probe readonly searches,/probe offkeeps Probe out.
Works with
Claude Code, Codex, Cursor (the app's agent and the cursor-agent CLI), pi and Kimi Code. Installing for one agent looks like this:
npx probe-research install --agent claude_code --yes