English | 中文

Sheaf

Harvest your knowledge. Bundle it. Share it.

Sheaf turns the links you save every day into a searchable knowledge base your AI agents can actually use. Paste a link — it fetches, classifies, summarizes. Crystallize many into portable knowledge cards. Local-first, open-source, no cloud.

A sheaf is a bundle of grain a farmer brings to market. Sheaf does the same for knowledge — gather, bundle, trade.

Quick Start

1 · Install (works everywhere):

pip install sheaf-ai
sheaf config setup     # one-time: pick a provider, paste an OpenAI-compatible API key

2 · Connect your agent:

sheaf setup            # auto-detects Claude Code / Codex / Cursor / Windsurf / WorkBuddy,
                       # writes the MCP config + deploys the bundled skill

3 · Use it:

sheaf collect https://arxiv.org/abs/2401.00000   # save a link
sheaf search "transformer architecture"          # search your collection
sheaf crystallize AI                             # distill knowledge cards

No accounts, no cloud. Your data lives locally — ./data/ inside a project, else ~/.sheaf/data — as Markdown + JSON. Override with SHEAF_DATA_DIR.

Even faster on Claude Code — no install at all:

claude mcp add sheaf -- uvx --from sheaf-ai sheaf-mcp

uvx is the npm-style runner (brew install uv / winget install astral-sh.uv). Then set a key via uvx --from sheaf-ai sheaf config setup.

Why

You save links every day — articles, papers, repos, tutorials. 90% never get opened again.

Not because you're lazy. Bookmarks serve human reading, not agent workflows. Ask your coding agent "what did I read about MCP last week?" — it has no idea.

Sheaf fixes this. Every link becomes a structured entry. Crystallize enough of them and you get knowledge cards — portable, searchable, agent-ready.

Features

Crystallize — your second brain

Sheaf's killer feature. Instead of letting bookmarks rot, sheaf crystallize synthesizes insights across entries:

$ sheaf crystallize AI
✨ 5 knowledge cards crystallized:
  📌 RAG faces retrieval relevance challenges (90%)
     RAG systems heavily depend on retrieval quality; errors degrade output reliability.
  📌 CRAG framework improves RAG robustness (95%)
     CRAG introduces a retrieval evaluator, web search augmentation, and document decomposition.

Each card carries a confidence score, evidence tracing (which sources contributed), topic provenance, and tags. Semantic-search across all of them with sheaf crystallize --semantic "query".

Connect Your Agent

sheaf setup writes the right MCP config for each tool and deploys a bundled skill / agents-note so the agent knows how to use Sheaf:

sheaf setup --target claude          # ~/.claude.json + ~/.claude/skills/sheaf-guide.md
sheaf setup --target codex           # ~/.codex/config.toml + ~/.codex/AGENTS.sheaf.md
sheaf setup --target cursor          # .cursor/mcp.json   (also: windsurf, workbuddy)
sheaf setup                          # auto-detect from CWD + installed agents
sheaf setup --target codex --dry-run # preview without writing

MCP + skill are one install. sheaf setup deploys the MCP server and the skill together — the skill is what tells your agent when to proactively capture a note or recall from the KB, so it's not optional decoration. Prefer it over the bare uvx one-liner (which wires MCP only). Do it all at once — key + MCP + skill + health check — with sheaf init --auto.

The MCP server exposes 4 core tools — sheaf_collect, sheaf_search, sheaf_crystallize, sheaf_get_card — covering ~90% of automated agent workflows and kept lean by design (~1.5k vs ~5k tokens). 7 more stay reachable via the sheaf CLI (--json) or MCP tools/call; re-expose all with SHEAF_MCP_TOOLS=all. Full tool matrix + rationale: Issue #91. Setup details: docs/mcp-setup.md.

Agents can also browse the knowledge base read-only via MCP Resources — sheaf://entries/recent, sheaf://entries/{id}, sheaf://stats, sheaf://tags (resources/list / resources/read). Spec: docs/agent-query-spec.md.

Commands

sheaf help                       # grouped command overview
sheaf collect <url> | --text "…" # save a link, or a pasted note (tagged 'note')
sheaf search <query>             # full-text search (results show the entry id)
sheaf list [--page N]            # browse entries, paginated
sheaf get <id>                   # full detail of one entry
sheaf crystallize <topic>        # crystallize knowledge cards from a topic
sheaf stats | tags | weekly | insights | urgent
sheaf mcp                        # start the MCP server (stdio)

Run sheaf <cmd> --help for per-command options.

Privacy & Local-First

Your data never leaves your machine unless you choose to.

• All content stored locally — ./data/ inside a project, otherwise ~/.sheaf/data (override: SHEAF_DATA_DIR)
• LLM calls go to your chosen API provider — nothing routed through Sheaf
• No telemetry, no analytics, no accounts
• Markdown + JSONL format — fully portable, zero lock-in

Configuration

sheaf config setup is the recommended path — interactive, OS-agnostic, stores keys in ~/.sheaf/config.json with restricted permissions. Any OpenAI-compatible endpoint works:

sheaf config use deepseek        # switch default provider
sheaf config list                # show configured providers

# macOS / Linux
export OPENAI_API_KEY=sk-...
# Windows PowerShell:   $env:OPENAI_API_KEY="sk-..."
# Windows CMD:          set OPENAI_API_KEY=sk-...
export OPENAI_BASE_URL=https://api.openai.com/v1   # optional — for non-OpenAI endpoints

Architecture

URL → fetch → classify → summarize → store → query
       (3-strategy   (LLM tags   (summary    (JSONL + MD
        fallback)     + topics)   + deadline)  index)
                     ↓
              crystallize → KnowledgeCard → EmbeddingEngine → semantic search

Requirements

• Python 3.10+
• An OpenAI-compatible API key
• Playwright Chromium (optional, JS-heavy sites): pip install -e ".[browser]" && playwright install chromium

Development

git clone https://github.com/zhelunSun/sheaf-ai.git && cd sheaf-ai
python -m pip install -e ".[dev]"
python -m pytest tests/ -q          # 1024 passed, 19 skipped
python -m ruff check sheaf_ai/ tests/ sheaf_cards/

Extras: .[dev] for local dev, .[server] for the HTTP API, .[browser] for Playwright fetching.

Status & Chrome Extension

Sheaf is early alpha. The collect → search → crystallize → MCP pipeline works and is covered by 1024 passing tests. We're validating with real users before beta.

A Chrome extension (extension/) adds one-click collect + search from any page: start the local API with sheaf serve, load extension/ unpacked (Chrome → Manage Extensions → Developer mode), then Alt+Shift+S or right-click any page → "🌾 Collect with Sheaf".

Try it: save 20+ links, run sheaf crystallize <topic>, then ask your agent to find them. If it clicks for you, open an issue or discussion and tell us what you'd change.

⭐ If Sheaf saves you time, a star on GitHub helps others find it.

License

Apache 2.0

A sheaf is a bundle of harvested grain — the unit a farmer brings to market. In mathematics, a sheaf attaches local data to open sets and glues them into a global picture. Sheaf the tool does both: gather scattered knowledge into coherent bundles, ready for your agents to consume or for you to share.