
claude-wiki
wiki: build, query and maintain a Markdown knowledge wiki from documents with Claude Code
Install with your AI
Paste into Claude Code, Cursor, or any agent — it reads the repo and wires the tool into your project.
Install and set up claude-wiki (git-clone project) into my current project. Found on https://claudeers.com/claude-wiki Repo: https://github.com/win4r/claude-wiki Homepage/docs: https://space.bilibili.com/3493277319825652 Detected install method: git-clone → git clone https://github.com/win4r/claude-wiki Category: mcp-servers. Platforms: cli, api, web. Read the repo's README for exact setup and env vars, then install it and wire it into my project. Claudeers Health Verdict: unknown; community-verified: false. Confirm the source before running anything.
git clone https://github.com/win4r/claude-wiki
// compatibility
| Platforms | cli, api, web |
|---|---|
| Operating systems | — |
| AI compatibility | claude |
| License | — |
| Pricing | open-source |
| Language | Python |
🚀观看我的哔哩哔哩视频
wiki
wiki has Claude Code build a Markdown knowledge wiki in a git repo from a folder of documents, answer
questions from it with a source tag on every fact, and keep it correct with reviewed dreaming passes. Every
agent step is one headless, restricted-mode claude -p run in its own git worktree; the program checks what each
run changed and records what it cost in a ledger. SPEC.md is the full design; DEVLOG.md says where things stand.
Setup
Requirements: Python 3.11 or later, git, and Claude Code 2.1.319 or later (claude --version).
pip install -e '.[convert]'
The convert extra adds python-docx, openpyxl, python-pptx and pypdf for Word, Excel, PowerPoint and
text-layer PDF files. Plain text, Markdown, HTML, CSV and email need nothing extra.
The API key and the Console spend limit
Paid runs need an API key. Each run uses claude --restricted --strict-mcp-config --disable-slash-commands
with a private, empty CLAUDE_CONFIG_DIR (runs/.claude-config), so your Claude login, settings, plugins, hooks,
MCP servers and auto memory never load, and CLAUDE.md files are switched off. (SPEC asks for --bare, but Claude
Code 2.1.319's bare mode only offers Bash, Edit and Read, so the owner approved this change on 2026-11-08; see
DEVLOG.md.)
- In the Claude Console, create a workspace for this tool and set its spend limit to $50. The tool's own caps rely on Claude Code's client-side cost estimates, which can differ from the bill (for example, Haiku 5.5 charges more for prompts over 100,000 tokens); the Console limit is the backstop.
- Create an API key in that workspace and export it before a paid command:
export ANTHROPIC_API_KEY=...(the tool never writes it to a file, log, commit or wiki page). - For unattended runs (cron), put an
apiKeyHelpercommand in a Claude Code settings file and pointrunner.api_key_settingsinwiki.tomlat it; the runner passes it with--settings:
{ "apiKeyHelper": "security find-generic-password -s wiki-anthropic-key -w" }
Quick start on the example corpus
wiki init ~/wikis/sec-deal
cd ~/wikis/sec-deal
export EDGAR_USER_AGENT="Your Name [email protected]" # SEC asks every client for a contact
wiki fetch-example --tier quickstart
wiki build --estimate
wiki build
wiki ask "What is the per-share merger consideration, and how did it change?"
Commands
Every command finds its workspace from --workspace <dir>, WIKI_WORKSPACE, or the nearest folder upward
holding wiki.toml. Commands that write the workspace or spend money hold <workspace>/.lock; a second one
exits with a message (exit code 75).
| Command | What it does | Paid |
|---|---|---|
wiki init <dir> | Creates a workspace: wiki.toml, a short CLAUDE.md, docs/, runs/ (mode 0700), ledger.jsonl, and repo/ (git) with a seeded wiki/_schema.md. --schema-dir picks another schema and steering | No |
wiki fetch-example --tier quickstart|core | Downloads the example SEC filings with the quickstart's fetch script, adds the synthetic Word analyst note and vendor CSV, runs wiki add on them, sets the workspace's tier (and build cap), and copies the question sets into repo/eval/. Refuses to run without EDGAR_USER_AGENT | No |
wiki add <paths> | Converts documents to text starting with a [SOURCE: ...] line, splits long ones into -part<n> units, stores them read-only in docs/, and records them in manifest.json. Refuses, before it writes anything, scanned PDFs, files it cannot read, and two files that make one slug. A document without a [SOURCE: ...] line of its own gets one with the day it was added, and keeps that day while its text is unchanged | No |
wiki build | Runs the eight phases (convert, batch, extract, check, resolve, link, compile, index) over new or changed units, committing after each phase, then wiki check. --resume finishes an interrupted build, rerunning only unfinished tasks | Yes |
wiki check | Lints the wiki and enforces its invariants; --against <ref> also checks that every per-unit fact line of that commit is still there, unchanged or with only valid_to/retracted added. Exits 1 on any error | No |
wiki ask "<question>" | One fresh read-only session; --mode wiki (the wiki only), sources (the documents only) or both (default: the wiki first, documents to confirm or fill a gap). In wiki and both modes the program hands the run a map of the wiki: the index's "Where to start" section, the page and other names (from _aliases.md) of each entity the question names, and the manifest's units in order of document date (a unit whose document gives none takes the date most of its facts carry, marked ~; of more than 80 units the latest are listed). A lowercase word of the question that is also a name ("block" for Block, Inc.) is offered as a maybe. The start section and the units, the same for every question, end the run's system prompt, which later runs read from the prompt cache; the names end its task. The run reads the Current picture of the one to three pages the question turns on, then searches the fact folders (notes/, metrics/, relations/, addenda/, contracts/) under every name, metrics/ first for figures, and the latest units for anything that can change. Checks every cited unit against the manifest, rewrites a tag that names a unit's file (metrics/<unit>.md) to name the unit, asks once more when an answer cites an id outside it, and saves the transcript to runs/ask/ | Yes |
wiki dream | With at least 5 ask transcripts since the last accepted dream: analysts find where readers struggled, one apply run confirms each finding in a document and files it, the program recompiles and reindexes, all on branch dream/<timestamp>, then wiki check --against its base. --accept merges it and tags dream-accepted-<timestamp>; --discard deletes it | Yes |
wiki eval <file> | Asks every question in every mode (--modes to narrow), grades each answer on Sonnet against its criteria, and writes repo/eval/report-*.md and .json: score and cost per mode, traps refused, and the break-even point | Yes |
wiki cost | The ledger: spent, remaining, spend by command and phase | No |
wiki status | Units by status, pages by kind, the last build and dream, any open review branch | No |
wiki smoke | M0's live checks of Claude Code (the key, init tools and plugins, Grep in an added folder, denied reads, no CLAUDE.md, a long-prompt cost). --probe-tools shows which tools each flag set starts | Yes, about $0.35 ($0.10 for the probe) |
Paid commands share --estimate (print the estimate and exit), --max-usd <n> (lower the command's cap),
--model <alias>, --concurrency <n> (default 4) and --approve-over-cap (the owner approves, once, an
estimate above the command's cap).
Costs
Every paid command prints its estimate first. The caps live in wiki.toml:
| Cap | Default |
|---|---|
| All paid runs, in total | $50 |
| One build, quickstart tier | $6 |
| One build, core tier | $25 |
| One question | $0.50 |
| One dream | $3 |
| One eval run | $3 |
- Build estimates start from the knowledge-wiki quickstart's figures, scaled by unit sizes; a question starts at $0.02, grading at $0.02 an answer, a dream at $1 plus $0.15 a transcript. After paid runs, the ledger's averages per phase and model replace them.
- A command's cap is split across its runs in proportion to their estimates, with 10% held back, and each run
gets its share as
--max-budget-usd. - An estimate above the command's cap stops with exit code 3 until the owner approves it once with
--approve-over-cap. If spent plus the estimate would pass the total, the command refuses (exit code 4) and names the shortfall; only the owner raisescaps.total_usd, by editingwiki.toml. - A run that ends without a result event is charged its full share.
- Ctrl-C, a
kill(SIGTERM) or a closed terminal (SIGHUP) stops a command's runs and starts no others. A run cut short is recorded in the ledger asinterrupted, charged the same way, and never retried;wiki build --resumeruns it again. ledger.jsonlis the only record of spend; it is outside git. Default models: Haiku for every agent role, Sonnet for the grader, set per role under[models].
Dreaming nightly
The tool never merges a dream on its own. To run one every night and leave the branch for review, add a cron
line like this (with runner.api_key_settings set, so the key comes from your apiKeyHelper):
15 3 * * * cd /Users/you/wikis/sec-deal && /usr/local/bin/wiki dream >> runs/dream-cron.log 2>&1
Then review with git -C repo diff <base>..dream/<timestamp> and run wiki dream --accept or --discard.
Accept or discard a dream before the next wiki build: if main has changed the same files since the dream
started, --accept refuses and leaves main as it was, and the dream has to be discarded and run again. A dream
that was stopped partway can only be discarded.
Whether to install it is your call.
Layout and safety
- The workspace's
docs/(normalized text) andruns/(transcripts, run inputs, worktrees, build state) are outside git; onlyrepo/is a git repo. Every run sends document text to the Anthropic API. - In
repo/wiki/, every file has one writer. Per-unit files (notes/,addenda/,contracts/,relations/,metrics/,open_questions/) are the source layer; the program compilesentities/<slug>.mdfrom them, and a Compile run may only change the text between a page's current-picture markers. - A writing run works in its own worktree and may change only the paths it owns; anything else fails the run
and discards the worktree. Inputs are read-only copies, re-hashed after every run and phase;
_schema.mdis verified unchanged after every phase. - No role gets Bash, web or MCP tools; the runner fails any run whose
system/initlists other tools than its role's (plusEndConversation), any MCP server or skill, or a plugin that isn't built in.
Development
pytest # offline: no network, no claude needed (a fake claude stands in)
WIKI_LIVE=1 pytest -m live # paid: spends real money through the ledger
Code adapted from anthropics/claude-quickstarts (MIT) keeps its notice in NOTICE.
// faq
What is claude-wiki?
wiki: build, query and maintain a Markdown knowledge wiki from documents with Claude Code. It is open-source on GitHub.
Is claude-wiki free to use?
claude-wiki is open-source, so it is free to use.
What category does claude-wiki belong to?
claude-wiki is listed under mcp-servers in the Claudeers registry of Claude-compatible tools.
// embed badge
[](https://claudeers.com/claude-wiki)
// retro hit counter
[](https://claudeers.com/claude-wiki)
// reviews
// guestbook
// related in MCP Servers
f.k.a. Awesome ChatGPT Prompts. Share, discover, and collect prompts from the community. Free and open source — self-host for your organization with complete…
A cross-platform desktop All-in-One assistant for Claude Code, Codex, OpenCode, OpenClaw, Gemini CLI & Hermes Agent. Only official website: ccswitch.io
🪨 why use many token when few token do trick — Claude Code skill that cuts 65% of tokens by talking like caveman
An open-source AI agent that brings the power of Gemini directly into your terminal.