Most AI agents still explore documentation the expensive way:
open file → skim hundreds of irrelevant paragraphs → open another file → repeat
That burns tokens, floods context windows with noise, and forces models to reason through a lot of text they never needed in the first place.
jDocMunch-MCP lets AI agents navigate documentation by section instead of reading files by brute force.
It indexes a documentation set once, then retrieves exactly the section the agent actually needs, with byte-precise extraction from the original file.
| Task | Traditional approach | With jDocMunch |
| --- | ---: | ---: |
| Find a configuration section | ~12,000 tokens | ~400 tokens |
| Browse documentation structure | ~40,000 tokens | ~800 tokens |
| Explore a full doc set | ~100,000 tokens | ~2,000 tokens |
Index once. Query cheaply forever. Precision context beats brute-force context.
jDocMunch MCP
AI-native documentation navigation for serious agents
⭐AI-driven public opinion & trend monitor with multi-platform aggregation, RSS, and smart alerts.🎯 告别信息过载，你的 AI 舆情监控助手与热点筛选工具！聚合多平台热点 + RSS 订阅，支持关键词精准筛选。AI 智能筛选新闻 + AI 翻译 + AI 分析简报直推手机，也支持接入 MCP 架构，赋能 AI 自然语言对话分析、情感洞察与趋势预测等。支持 Docker ，数据本地/云端自持。集成微信/飞书/钉钉/Telegram/邮件/ntfy/bark/slack 等渠道智能推送。
Every 1.x license entitles you to every future 1.x release. We will never ship a 1.x version that:
removes or renames an MCP tool (deprecated tool names keep their aliases),
drops a Section field from the response shape,
forces a reindex without auto-migrating your existing index on first load,
changes the JSON wire format of any tool response in a way that breaks an existing consumer,
or makes a previously-default behavior raise.
Anything that would require breaking these promises is reserved for a future major version (2.x). The full machine-checked contract is enforced via tests/test_server.py (tool-name and required-field invariants) and the replay-fixture gate that runs on every release.
jDocMunch indexes documentation once by heading hierarchy and section structure, then gives MCP-compatible agents precise access to the explanations they actually need instead of forcing them to brute-read files.
It is built for workflows where token efficiency, context hygiene, and agent reliability matter.
Why this exists
Large context windows do not fix bad retrieval.
Agents waste money and reasoning bandwidth when they:
open entire documents to find one configuration block
repeatedly re-read headings, boilerplate, and unrelated sections
lose important explanations inside oversized context payloads
consume documentation as flat text instead of structured knowledge
jDocMunch fixes that by changing the unit of access from file to section.
Instead of handing an agent an entire document, it can retrieve exactly:
an installation section
a configuration section
an API explanation
a troubleshooting section
a specific subtree of related headings
That makes documentation exploration cheaper, faster, and more stable.
What makes it different
Section-first retrieval
Search and retrieve documentation by section, not just file path or keyword match.
Byte-precise extraction
Full content is pulled on demand from exact byte offsets into the original file.
Stable section IDs
Sections retain durable identities across re-indexing when path, heading text, and heading level remain unchanged.
Local-first architecture
Indexes and raw docs are stored locally. No hosted dependency required.
MCP-native workflow
Works with Claude Desktop, Claude Code, Google Antigravity, and other MCP-compatible clients.
What gets indexed
Every section stores:
title and heading level
one-line summary
extracted tags and references
SHA-256 content hash for drift detection
byte offsets into the original file
This allows agents to discover documentation structurally, then request only the specific section they need.
Why agents need this
Traditional doc retrieval methods all break in different ways:
File scanning loads far too much irrelevant text
Keyword search finds terms but often loses context
Chunking breaks authored hierarchy and separates explanations from examples
jDocMunch preserves the structure the human author intended:
heading hierarchy
parent/child relationships
section boundaries
coherent explanatory units
Agents do not need bigger context windows.
They need better navigation.
How it works
jDocMunch implements jMRI-Full — the open specification for structured retrieval MCP servers. jMRI-Full covers the full stack: discover, search, retrieve, and metadata operations with batch retrieval, hash-based drift detection, byte-offset addressing, and a complete _meta envelope on every call.
Parsing
Format-aware section splitting: heading-based (Markdown/MDX/HTML/RST/AsciiDoc), structure-based (OpenAPI tags, JSON keys, XML elements), or cell-based (Jupyter)
Hierarchy wiring
Parent/child relationships established
Summarization
Heading text → AI batch summaries → title fallback
Storage
JSON index + raw files stored locally under ~/.doc-index/
Retrieval
O(1) byte-offset seeking via stable section IDs
Stable section IDs
{repo}::{doc_path}::{ancestor-chain/slug}#{level}
The slug is prefixed with the ancestor heading chain, making IDs both readable and stable. A new heading inserted in one branch of a document never renumbers IDs in another branch.
IDs remain stable across re-indexing when the file path, heading text, heading level, and parent heading chain do not change.
Installation
Prerequisites
Python 3.10+
pip
Install
pip install jdocmunch-mcp
Verify:
jdocmunch-mcp --help
Configure an MCP client
PATH note: MCP clients often run with a restricted environment where jdocmunch-mcp may not be found even if it works in your shell. Using uvx is the recommended approach because it resolves the package on demand without relying on your system PATH. If you prefer pip install, use the absolute path to the executable instead.
For Anthropic or Gemini, the base uvx jdocmunch-mcp command is enough once the
corresponding API key is present. For OpenAI-compatible providers such as OpenAI,
MiniMax, or GLM-5, include the optional dependency in the launcher command:
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
A Claude Code plugin that automatically captures everything Claude does during your coding sessions, compresses it with AI (using Claude's agent-sdk), and injects relevant context back into future sessions.
