Track your autonomous research and experimentation, send follow-up feedbacks, and analyze reports and insights with autoexp — a local, browser-based experimentation surface for AI coding agents: Claude Code, Codex, OpenCode, Pi.
AI agents are already good at proposing code. The harder problem is the harness around those proposals: source boundaries, reproducible execution, external inputs, evaluators, artifacts, lineage, rollback, and human review. Autoexp supplies that infrastructure while your repository remains the editable source of truth.
Explore the workflow and installation options at autoexp.dev.
|
Compare variants, inspect immutable evidence, and turn results into a clear recommendation. |
|
|
Optimize a measurable objective with a frozen evaluator and a keep-or-revert loop. |
|
macOS, Linux, and WSL (requires Git, curl, and uv):
curl -fsSL https://autoexp.dev/install.sh | bashWindows PowerShell (requires Git and uv):
irm https://autoexp.dev/install.ps1 | iexRestart your coding agent, open a repository, and start with an objective.
Autoexp exposes two agent workflows:
| Workflow | Codex | Claude Code | OpenCode | Pi |
|---|---|---|---|---|
| Start or continue experiments | $autoexp <objective> |
/autoexp <objective> |
/autoexp <objective> |
/autoexp <objective> |
| Open browser feedback review | $autoexp-review |
/autoexp-review |
/autoexp-review |
/autoexp-review |
/autoexp Compare the cache strategies in this repository. Reuse the existing
replay benchmark, preserve every run, and recommend a winner from the evidence.
Autoexp understands the objective, works with the code, data, and evaluators already in your repository, and preserves every result.
You define the objective and success criteria
↓
Your agent proposes a focused repository change
↓
Autoexp pins execution and seals the resulting evidence
↓
A metric or human review decides what happens next
Invoke the review workflow when you want to inspect results or steer the next step:
/autoexp-review
The review command opens a local browser session immediately. You can inspect source, rendered artifacts, CSV tables, images, logs, reports, and diffs. Add notes and submit one feedback batch. Your agent receives it and continues the experiment.
Use autoexp view to browse and download experiments, runs, artifacts, reports,
metrics, and diffs in one place.
The local runner executes trusted experiment code as your operating-system user; it is not a sandbox. Use the Docker runner for generated or untrusted programs. Docker runs drop capabilities, prevent privilege escalation, use read-only source/run mounts, and expose only output/report directories as writable, but they may still have network access if enabled and are not a hardened multi-tenant sandbox.
Runs receive only declared, file-backed, or explicitly supplied environment values. Autoexp redacts known raw secret values from recorded evidence, but cannot recognize encoded, hashed, split, or transformed secrets. Use short-lived, least-privilege credentials and review evidence before publishing it.
Use these commands for history, automation, and direct control:
| Task | Command |
|---|---|
| Show registered experiments | autoexp experiment list |
| Inspect recent runs | autoexp status |
| Open the global dashboard | autoexp view |
| Compare two immutable runs | autoexp diff <run-a> <run-b> |
| Restore declared source from a run | autoexp restore <run-id> |
| Check the selected experiment and runtime | autoexp doctor |
| Open a blocking agent review | autoexp review |
Autoexp stores experiment history and restorable snapshots locally. Repository files stay in their original locations. Declared secret-source files are excluded from snapshots and browser reads. Review screens display recorded outputs, so keep sensitive values outside experiment outputs.
Default data directory:
- Linux:
$XDG_DATA_HOME/autoexpor~/.local/share/autoexp - macOS:
~/Library/Application Support/autoexp - Windows:
%LOCALAPPDATA%/autoexp
Set AUTOEXP_HOME to override it. Use autoexp relink <repo-id> <new-path> if a worktree moves.