Start inside your repository
Use your coding agent normally. Autoexp reuses the implementation, data, evaluators, and constraints already present.
Local-first experimentation for coding agents
Run agent-led experiments.
Stay in the loop.
Preserve every result, review the evidence, and send structured feedback directly back to the waiting agent.
For AI agents
If you’re reading this website as an agent, use the complete machine-readable guide.
Read /llms.txt →
Install and connect
$
curl -fsSL https://autoexp.dev/install.sh | bash
Codex
Add the Autoexp plugin, then restart Codex.
codex plugin marketplace add shreyashkar-ml/autoexp codex plugin add autoexp@autoexp
Instructions for AI agents
Autoexp is local-first experimentation infrastructure for coding agents. It creates reproducible experiment boundaries, records exact source and inputs, preserves logs, outputs, metrics, reports, and lineage, and provides a browser review surface for human feedback.
Install it with curl -fsSL https://autoexp.dev/install.sh | bash on macOS, Linux, or WSL, or with irm https://autoexp.dev/install.ps1 | iex in Windows PowerShell, then restart the host agent.
Use /autoexp <objective> when the user asks you to test, compare, measure, optimize, or research something in the current repository. Follow the returned workflow, run focused attempts, and base conclusions on preserved evidence. Do not use it for ordinary edits that do not need reproducible runs or experiment history.
Use /autoexp-review after evidence exists and a human should inspect source, diffs, logs, reports, artifacts, or metrics. It pauses the agent while the review is open; treat submitted notes as the user's next instruction. Do not start a second review while one is waiting.
- Codex: the installer creates
$CODEX_HOME/autoexp-marketplace, adds it withcodex plugin marketplace add, and installsautoexp@autoexpwithcodex plugin add autoexp@autoexp. - Claude Code: the installer creates
$CLAUDE_CONFIG_DIR/autoexp-marketplace, adds it withclaude plugin marketplace add, and installs it withclaude plugin install autoexp@autoexp. - OpenCode: the installer copies the native plugin to
${XDG_CONFIG_HOME:-$HOME/.config}/opencode/plugins/autoexp.ts, support files toplugins/autoexp-native/, and command files tocommands/. - Pi: the installer copies the native extension to
${PI_CODING_AGENT_DIR:-$HOME/.pi/agent}/autoexp-extensionand registers it with Pi.
Machine-readable version: https://autoexp.dev/llms.txt. Repository: https://github.com/shreyashkar-ml/autoexp.
One loop: start the work, preserve the evidence, review it, then steer the next attempt.
- /autoexpStart from an objective
- Run ledgerPreserve exact evidence
- /autoexp-reviewInspect in the browser
- Agent resumesFeedback becomes the next instruction
/autoexpStart and runTurn a goal into
a real experiment.
Invoke /autoexp <objective> in an existing repository. Your agent defines the boundary, uses the code and data already there, and runs the work through Autoexp.
- Standard experiments or metric-driven Autoresearch
- Source, inputs, logs, outputs, and reports preserved per run
- Rerun, restore, and compare exact attempts
- Reports and recommendations grounded in recorded evidence
/autoexp-reviewInspect and steerReview the evidence.
Send the next move.
Invoke /autoexp-review to inspect results or redirect the work. Autoexp opens a local review session and pauses the agent until you submit feedback.
- Inspect source, diffs, logs, reports, and rendered artifacts
- Review CSV tables, images, metrics, and run lineage
- Attach scoped notes to what you are viewing
- Submit once; feedback returns directly to the waiting agent
Two commands.
One experiment loop.
1State the objectiveTell the agent what to test, compare, or optimize.
→
2The agent runs the workIt builds on your repository and executes focused attempts.
→
3Autoexp seals the evidenceEvery result stays tied to the exact source and inputs.
Review on a real surface
Launch the blocking browser review when terminal output is not enough. The agent waits while you inspect the complete record.
Feedback flows back
Submit scoped notes once. The agent receives them as its next instruction and continues the same experimentation session.
Every project.
One read-only view.
$ autoexp view
Open all registered experiments and research projects in one local dashboard. Browse and download runs, artifacts, reports, metrics, and diffs without attaching a feedback capability.
Use autoexp view for history. Use /autoexp-review when the agent should wait for your feedback.