Skip to content

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

anchor serve

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:

anchor ingest /path/to/datasheet.pdf
anchor list

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.

Built-in ingest vs. harness-driven ingest

ANCHOR has two ways to turn a PDF into usable canvas evidence:

Mode Entry points Gold extraction path Best fit
Built-in ingest Canvas drag-drop, HTTP upload, MCP ingest_pdf, CLI anchor ingest ANCHOR runs Docling silver extraction and, when configured, a vision extractor for gold regions Fast normal use and repeatable scripted ingestion
Harness-driven ingest MCP ingest_begin, ingest_get_page, ingest_submit_page, ingest_finalize The connected agent reads each page work item and submits polished markdown plus regions No-key provider, quality-sensitive tables, or dense datasheets

Both modes write the same bronze, silver, and gold folders. The difference is who performs the gold-region interpretation. Drag-drop and anchor ingest use the configured backend extractor. Harness-driven ingest lets the agent review each page before publishing gold, so it can be slower but more deliberate on difficult tables.

3. Create a canvas

anchor canvas create pump-analysis --title "Pump analysis"
anchor canvas state pump-analysis

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:

anchor install claude-code   # MCP entry + skill (default env)
anchor install cursor

These register a server that resolves the project from the folder you open the agent in, with no baked data dir and 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.