Sheaf · Manual
A CRM in markdown.
The longer story, for the curious — what Sheaf is, what it does on your behalf, and what it doesn’t try to do. If your AI sent you here and you just want to start, the homepage’s three on-ramps will get you set up in under two minutes.
The model.
Your CRM is a folder. Every contact is a markdown file. Every deal is a markdown file. Every interaction — meeting, call, email, note — is a markdown file. Each file has structured frontmatter (name, email, last_contact, status, amount, expected_close) and a freeform body where the actual story lives (what you talked about, what to remember, what to follow up on).
Files reference each other through wiki-links: [[people/alex-chen]]in a deal’s body becomes a clickable graph edge.
The agent.
On import — Salesforce, HubSpot, Raiser’s Edge, or a spreadsheet — Claude Opus 4.7 reads your source schema and proposes a clean folder shape: which entities exist, what their frontmatter looks like, how they relate. You review and edit the proposal in plain English. Then the agent renders one markdown file per record.
On every read and write, Claude (in claude.ai over MCP) opens files, follows wiki-links, edits frontmatter, and logs interactions. When you say “Coffee with Alex today, he wants to expand from 50 to 200 seats,” Claude writes a new interactions/2026-MM-DD-acme-coffee.md file and updates deals/q3-renewal-acme.md.
The two surfaces.
- File ops — read_entity, write_entity, find, link, recent. What Claude uses on every conversation. Native to its Read / Edit / Glob primitives.
run_sql— a derived Postgres index of every file, for the questions only aggregation can answer. Top 10 deals, donor cohorts, pipeline by stage.
What Sheaf charges for.
- The import.Turning 5,000 Salesforce records into well-shaped markdown — with inferred frontmatter, deduplication, wiki-links — is real agent work. Done by hand it’s 20+ hours.
- The hosted MCP layer. Secure OAuth, DCR, bearer auth — what lets claude.ai actually talk to your folder.
- The Postgres index. So
run_sqlworks fast over thousands of files. Plus the indexer that keeps it in sync as you edit files. - Sync maintenance. OAuth refresh, schema drift handling, cron — the boring plumbing that keeps your CRM flowing in.
- Curated tools. Per-tenant agent-generated MCP tools beyond read/write — like
find_lapsing_donorsordeals_at_risk_q3.
What Sheaf doesn’t do.
- It isn’t a dashboard.There’s no Sheaf UI for browsing your CRM in a table. Claude is the UI; your editor is the inspection tool.
- It doesn’t hold your CRM’s system of record. v1 mirrors SF / HubSpot / RE read-only. Your existing CRM stays authoritative. Write-back is a v2 question.
- It doesn’t train models on your data. Per-call requests to Claude only — no fine-tunes, no aggregation, no shared corpus.
Files you can hold.
Your folder is yours. Export it as a tarball any time. Mirror it into a private git repo for full version history. Open any file in your editor and read exactly what Claude reads — there’s no hidden binary, no proprietary format, no opaque DB.