Agent-first document management
Your agent works on your cases directly: Claude through the MCP connector, any other agent through the REST API.
What an outside agent can do
… stands for /api/applications/{id}.
| Capability | Example | Claude (connector tool) | Any agent (API call) |
|---|---|---|---|
| Find and open cases | Cases by how complete their document list is | list_applications, lookup_application | GET /api/applications, GET … |
| Create a case | A case on the Purchase Loan (CA) list | list_expected_doc_lists, create_application | POST /api/applications |
| Add files | Send two PDFs from Claude’s sandbox | create_upload_link | POST …/files |
| Request documents from a customer | Ask the borrower for pay stubs and bank statements | create_request_link | POST …/request-links |
| Split, classify, extract | Run the case’s workflow on new files | list_workflows, start_workflow_run | POST /api/workflows/{id}/runs, POST …/run/gate, POST …/run/cancel |
| Search and read pages | Find “loan amount”, then read that page | search_case_text, read_parsed_pages | GET …/search, GET …/text, GET …/pages-pdf |
| List and download documents | A filed document’s PDF | list_documents | GET …/documents, GET …/approved-docs/{docId}/file |
| Read extracted values | Net pay from each pay stub | — | GET …/analysis, POST …/run/tool |
| Match a document | Pages 12–13 are the W-2 | approve_document | POST …/document-approvals |
| File pages no document claimed | File page 40 under Gift Letter, or archive it | file_pages | POST …/placements |
| Tag a document | Priority, or Excluded from search | tag_document | POST …/document-tags |
| Cited briefs | A cited summary of the case | list_intel_templates, start_intel_brief, list_intel_briefs, get_intel_brief | POST …/intel-briefs, GET …/intel-briefs/{briefId} |
| Teach (fine-tune) | Add a test for a misread total, then read the report | — | POST …/files/{fileId}/pages/{page}/teach-tests, GET …/teach-report |
| Audit trail | Who changed what, and when | — | GET …/events |
| Webhooks | A signed POST when a run finishes | — | a Manager adds the endpoint |
| Stored files | Your folders and files, with sizes and dates | list_storage | GET /api/storage |
| Models | The model options for each step | — | GET /api/models/options |
How an agent connects
- Claude (claude.ai, the Claude app, Cowork, Claude Code): add the connector
https://sheaf.us/api/mcpand sign in once. It acts as that person, with their role. Connect Claude - Any agent that runs code (Claude Code, Codex): paste the setup from Settings → API keys, or connect with a pairing code. An API key acts as an Originator in the organization. Connect an agent
- Claude, ready-made: the skill or the plugin,
sheaf-skill.ziporsheaf-claude-plugin.zip. The plugin adds the connector too. Add the skill - Try it:
POST /api/guestreturns a key for a guest sandbox with a sample case, open for 24 hours, at 10 requests a minute. - Docs an agent reads:
/skill.md,/llms.txt, any docs page as.md, and/api/openapi.jsonwithout a key. Docs for agents
Your data
- Your documents can sit in your own S3 bucket in a standard folder layout, so a project or the whole library copies out. You keep your data
- Your database, your storage, your model keys. Whitelabel
- You keep all rights, with 30 days to export after you leave. Terms
- Role checks on every request, and fixed retention. Security and data
Safeguards
- Every change an agent makes is in the audit trail, shown, for example, as “Agent via Claude · jane@example.com”.
- An agent can match documents but never reject them. Only a person gives Teach verdicts.
- Flagged files wait for a person. Another organization’s case answers “not found”.
Sheaf