docagent-ai 1.0.4

Whole-repo documentation generator (README, AGENTS.md, CLAUDE.md, how-to guides, API reference, llms.txt) with deterministic verifier and citation grounding. Not affiliated with facebookresearch/DocAgent (Meta AI, arXiv 2504.08725).

docagent

Repository documentation agent for humans and coding agents.

Not affiliated with facebookresearch/DocAgent (Meta AI, arXiv 2504.08725, ACL 2025). That project is a multi-agent pipeline for generating Python docstrings. This project is a single-agent CLI that generates whole-repository documentation (README, AGENTS.md, CLAUDE.md, how-to guides, API reference, llms.txt) with a deterministic verifier and citation-grounded output. See How we differ below.

Why

Repositories accumulate stale READMEs, missing how-to guides, and undocumented APIs faster than humans can maintain them, and AI coding assistants increasingly need their own orientation files (AGENTS.md, CLAUDE.md, llms.txt) alongside the human-facing docs. DocAgent treats documentation as a set of verifiable artifacts driven by a DAG, generates them with an LLM backend, and verifies every claim against the actual source via a deterministic-first pipeline. The verifier is built around ground-citation comments (<!-- ground: path:start-end -->) so generated docs stay anchored to real code.

Install

DocAgent targets Python 3.11+ and is distributed on PyPI as the docagent-ai package (the docagent name itself is blocked by PyPI's similarity policy due to neighbours like doc-agent and docs-agent; the imported module and the CLI binary remain docagent).

pip install docagent-ai

For the multi-provider backend (Gemini, OpenRouter, Anthropic-direct via LiteLLM), install the optional multi extra:

pip install 'docagent-ai[multi]'

Provider setup

DocAgent ships two backends. The default is the Claude Agent SDK (agent_sdk), which delegates to your local claude CLI.

# Default backend — uses your existing `claude` CLI session.
docagent init

The opt-in LiteLLM backend routes to Gemini, OpenRouter, Anthropic-direct, or OpenAI based on the --model string and requires the corresponding provider API key:

Provider	Model string example	Env var
Anthropic (direct)	anthropic/claude-sonnet-4-6	ANTHROPIC_API_KEY
Google Gemini	gemini/gemini-2.5-flash, gemini/gemini-2.5-pro	GEMINI_API_KEY
OpenRouter	openrouter/anthropic/claude-sonnet-4-6	OPENROUTER_API_KEY

export GEMINI_API_KEY=...
docagent init --backend litellm --model gemini/gemini-2.5-pro

--backend litellm without --model exits with a multi-line hint listing the supported routing strings.

Quickstart

Installation registers a single docagent console entry point. The Typer app exposes three commands: init, update, and verify.

# Full pass: scan repo, build symbol index, generate all artifacts.
docagent init

# Incremental refresh of artifacts affected by recent changes.
docagent update

# Re-run the deterministic-first verifier against artifacts on disk.
docagent verify

Useful flags on init/update: --dry-run previews diffs without writing, --only <artifact_id> restricts the run to specific artifacts, and --backend litellm --model <provider/model> swaps the default Claude Agent SDK backend for LiteLLM.

Architecture

DocAgent is organized into orthogonal packages under docagent/:

• docagent.cli — Typer-based entry point for init / update / verify; selects a backend and constructs the orchestrator.
• docagent.artifacts — Each artifact (readme, api_reference, how_to_guides, agents_md, claude_md, llms_txt, python_docstrings) owns its own plan → generate → verify cycle and registers into the DAG via register_v1_builtins.
• docagent.core — Orchestrator, scanner, diff/state tracking, and the BudgetTracker drive the DAG; CLI imports these directly.
• docagent.adapters + docagent.parser — Language adapters using libcst/jedi for Python and tree-sitter for Rust/Go/TypeScript/Java/C++ extract symbols and signatures.
• docagent.index — SQLite-backed symbol/mention store opened via open_store; populated by the scanner during init.
• docagent.backends — Pluggable LLM backends: agent_sdk (default, Claude Agent SDK) and litellm (multi-provider; --model required).
• docagent.verify — Deterministic-first verification pipeline backing the verify command and the per-artifact verify phase.

Status

Beta. The artifact pipeline runs end-to-end and self-verifies through the deterministic gates; the package version is 1.0.4 and APIs may still change between minor releases.

How we differ

This project shares the name "DocAgent" with Meta AI's facebookresearch/DocAgent (arXiv 2504.08725, ACL 2025 demo) but is a separate, unaffiliated project with a different scope:

Axis	Meta's DocAgent	This project
Scope	Python docstrings only, symbol-level	Whole-repo artifacts: README, AGENTS.md, CLAUDE.md, how-to guides, API reference, llms.txt, plus docstrings
Topology	5-agent pipeline (Navigator → Reader / Searcher / Writer / Verifier / Orchestrator)	Single-agent orchestrator over the Claude Agent SDK
Verification	LLM "Verifier" agent	Deterministic-first gate pipeline (ground-citation validation, markdownlint, structure, secrets)
Indexing	In-process dependency graph	Persisted SQLite symbol index (libcst + tree-sitter)
Grounding	Hierarchical context build	<!-- ground: path:line-start-line-end --> citation enforcement on every non-trivial claim
Distribution	Research repo, clone-and-run	pip install-able CLI: docagent init / update / verify

If you arrived here looking for the Meta paper's reference implementation, you want facebookresearch/DocAgent.

License

MIT.