agentlantern 0.1.22

Devtool to document and inspect LLM agent systems.

AgentLantern

GitHub: https://github.com/brellsanwouo/agentlantern

AgentLantern is a CLI devtool that automatically documents multi-agent LLM systems and lets you watch them run live.

Point it at an agent project — it detects the framework, parses the architecture, and generates a browsable documentation site. For supported frameworks it also opens a live animated playground while the crew executes.

Install

With pip:

pip install agentlantern

With uv (recommended — installs the lantern command globally):

uv tool install agentlantern

From source:

git clone https://github.com/brellsanwouo/agentlantern
cd agentlantern
uv tool install -e .
# or: pip install -e .

Requires Python 3.11+.

Quick Start

# Generate and browse documentation for your agent project
lantern web path/to/your-project

# Watch a CrewAI run live in the animated playground
lantern play path/to/your-project

# Generate docs only (written to <project>/docs/)
lantern docs path/to/your-project

# Lint the agent project for design issues
lantern lint path/to/your-project

Commands

Command	Description
lantern docs [project] [-o DIR]	Generate Markdown documentation
lantern web [project] [--port PORT]	Generate and serve project docs over local HTTP on 0.0.0.0:9000 by default
lantern lint [project] [--strict] [--json]	Static analysis — no LLM, no network
lantern inspect [project]	Output project model as JSON
lantern play [project] [--name NAME]	Live animated playground; if --name is omitted, the UI asks for a run name
lantern replay NAME [--speed N]	Replay a saved run; use last for the newest saved replay

Supported Frameworks

Framework	Status
CrewAI	Full analysis and live playground
LangGraph	Detected — analysis coming soon
AutoGen	Detected — analysis coming soon
Smolagents	Detected — analysis coming soon
Google ADK	Detected — analysis coming soon

Detection is automatic — no configuration needed.

What Gets Generated

lantern docs writes these files into <project>/docs/:

File	Content
overview.md	Project snapshot, entrypoints, high-level flow
architecture.md	System map, key files, dependencies, env vars
agents.md	Agent roles, goals, tools, backstories
tasks.md	Task descriptions, agents, expected outputs
diagrams.md	Mermaid diagrams: agent-task graph, execution flow, sequence
runbook.md	Install, configure, run, static checks, troubleshooting
contact.md	Project contacts from pyproject.toml
index.html	Docsify site (publishable to GitHub Pages)
agentlantern-docs.html	Self-contained shareable HTML bundle

Play UI

lantern play opens a pixel-art playground that visualizes a supported agent run:

• START / STOP / RESTART controls
• Run name field in the UI when --name is not provided
• Up to 10 agents, each in their own named zone on the map
• A central Tool Hub agents visually consult when tools are used
• Timeline, Thoughts, Tools, Comms, and Log panels
• Clickable agents with full per-agent history
• Final report display on completion
• Replay — save a run with --name my-run, replay it with lantern replay my-run, or open the newest saved run with lantern replay last

Linter

lantern lint analyzes your project statically:

AgentLantern Lint — my-crew (CrewAI)

  ✗  [E002]  crew.py:18   Task `analyze` references agent `researcher` which is not declared in agents.yaml.
  △  [W006]  agents.yaml:4   Agent `writer` is missing required field `backstory`.
  ·  [I001]  No tool detected — may be intentional for a pure LLM workflow.

  Summary: 1 error, 1 warning, 1 info

Rules: E001–E002 (errors), W001–W008 (warnings), I001–I004 (info).

Contributing

src/agentlantern/
├── core/            # BaseProject, BaseAnalyzer, BaseRenderer
├── frameworks/
│   ├── registry.py  # auto-detection
│   ├── crewai/      # full analyzer + renderer
│   ├── langgraph/   # stub
│   ├── autogen/     # stub
│   ├── smolagents/  # stub
│   └── google_adk/  # stub
├── docs.py          # doc orchestrator
├── play.py          # live playground server
└── cli.py

To add a new framework: create src/agentlantern/frameworks/<name>/analyzer.py implementing detect, analyze, get_renderer, then register it in registry.py.

License

MIT — see LICENSE