Working with documents and canvases¶
ANCHOR keeps documents and canvases separate: ingest a source document once, then reuse it on one or more canvases while retaining page and region provenance.
1. Start ANCHOR¶
Open http://127.0.0.1:8002. The HTTP server is unauthenticated and binds to loopback by default.
2. Ingest a document¶
Use the canvas upload surface, or ingest a PDF from the command line:
Bronze data stores the original input, silver data contains page extraction, and gold data contains structured regions when an LLM-backed extractor is configured.
Each completed ingest writes a timing report at
<data-dir>/silver/<slug>/ingest-report.json. The report records total
duration, stage duration, per-page polish timing, per-page gold extraction
timing, and embedding time. Use it when comparing slow and fast runs.
3. Create a canvas¶
anchor canvas create pump-analysis --title "Pump analysis" --data-dir ~/anchor-data
anchor canvas state pump-analysis --data-dir ~/anchor-data
Canvases are event-backed workspaces. Nodes may contain source_ref values so
facts remain linked to document pages or extracted regions.
4. Connect an agent¶
For supported MCP-capable clients:
These register a server that resolves the project from the folder you open the agent in — no baked data dir, no reinstall per project.
Restart the client after installation. An agent can enumerate workspaces, search ingested documents, add evidence-backed nodes, and organize a canvas through the MCP tools.
See Agent configuration for Codex, OpenCode, and generic stdio examples.
5. Work without an LLM key¶
ANCHOR can store PDFs, render pages and manage canvases without an external LLM. Configure an OpenAI-compatible vision endpoint only when you need gold-region extraction and grounded semantic lookup.
See Configuration for supported settings and MCP tools for agent-facing capabilities.