the_lab.api

Autonomous experiment management for AI research agents.
Let agents run experiments, branch ideas, and track metrics — hands-free.

The Lab is a lightweight API + dashboard that lets AI agents run structured research. No database, no setup beyond installing the CLI and running the-lab init.

Jump To: Quick Start, How It Works, Under the Hood

---

Quick Start

# User install
pipx install git+https://github.com/LambdaLabsML/the_lab.api.git

# Or install from source while developing
git clone https://github.com/LambdaLabsML/the_lab.api.git
cd the_lab.api
pip install -e .

# Initialize your project (interactive — sets up prompts, MCP, gitignore)
cd /path/to/your/repo
the-lab init

# Start the API + dashboard
the-lab .

# Launch one agent turn, or keep it looping
the-lab-agent
the-lab-agent loop -d 15m

That's it. Open http://localhost:8000 and watch the agent work.

See LambdaLabsML/the_lab.api.demo for a working example project.

the-lab . --port 9009           # bind a different port
the-lab . --host 127.0.0.1      # bind a different host
the-lab . --https               # self-signed HTTPS
the-lab . --perf                # write .the_lab/api_perf.csv
the-lab-agent --role instructor # use .the_lab/PROMPT.instructor.md
the-lab-agent --sandbox on      # enable network + file sandbox
the-lab-agent --agent codex     # Codex instead of Claude

To enable HTTP Basic Auth, set credentials before starting the server — the agent and MCP bridge pick them up automatically from the same env vars:

export THE_LAB_USER=alice
export THE_LAB_PASSWORD=secret
the-lab .                  # server enforces auth
the-lab-agent              # agent authenticates automatically

---

How It Works

1. the-lab init scaffolds .the_lab/PROMPT.md and optionally has Claude analyze your repo and fill it in — describe your research goal, background, and setup there
2. the-lab-agent registers an isolated per-agent worktree, then launches Claude or Codex with the MCP bridge attached
3. The agent calls get_instructions() to fetch the current prompt + API reference, creates ideas as git branches, runs experiments, and iterates
4. The dashboard shows progress in real time: metrics, graphs, tables, experiment logs, rendered output, agent state, prompts, sandbox config, and suggestions

Ideas form a DAG — each idea is a git branch that can fork from parents. Experiments and agents each run in isolated worktrees, so concurrent work does not collide.

Panes — the dashboard is a dockable layout; every pane can be floated, resized, or hidden to the tray:

Pane	What it shows / does
Filters	Global filter bar — by tag, status, experiment mode; affects all charts and table
Metrics	Line chart of any logged metric over time — log scale, outlier clipping, best-line overlay, colorblind palette
Scatter	Cross-metric scatter — plot any two metrics against each other
Table	Sortable multi-metric comparison across all experiments
Graph	Idea DAG — relationships, status, and improvement at a glance
Timeline	Experiments and notes interleaved chronologically
Log	Live experiment log viewer with full-text search
Detail	Per-idea drill-down: experiments, notes, log, output, script, diff
Agents	Active and past agent sessions — output log, token cost, resume command
Queue	Running and queued experiments; priority and resource assignment
Messages	Inter-agent messages and notifications
Sandbox	Network allow/deny lists + per-path file rules (rw / ro / hidden)
Prompts	Edit role-specific PROMPT.<role>.md files; copy agent launch command
Task	Current task description and status
Stats	API request timing and sizes
Suggest	Submit a new idea directly from the dashboard
API	Inline API reference

---

Under the Hood

• No database — all state in .the_lab/ as JSON files (gitignored, created automatically)
• Ideas are git branches — checkout, uncommitted changes carry to the new branch automatically
• Isolated worktrees — each experiment and each agent gets its own git worktree; concurrent runs never collide
• Survives restarts — running experiments re-attach on server startup
• MCP + curl — agents use typed MCP tools or plain HTTP, both work
• Sandbox — rootlesskit + bubblewrap + transparent proxy; default-deny network with per-path file rules (rw / ro / hidden)
• Auth — optional Basic Auth; scoped bearer tokens for experiment callbacks