ANCHOR¶
Agent-Native Canvas to Help Organize Resources.
ANCHOR is a local canvas where you and an external AI agent work with engineering documents. Add a PDF, ask for one specification table, and inspect each row's source in the PDF dock. Requests, clarifications, and proposed edits can stay attached to the canvas objects they concern.
The browser, CLI, and MCP tools use the same persisted project data. ANCHOR runs on your computer; document disclosure depends on the provider and agent you choose. A local project folder does not imply that the connected model is local.
Start here¶
- Install ANCHOR. The published wheel includes the browser UI; Python 3.12+ is required.
- Follow the Quickstart. Choose a provider, create a project and canvas, connect an agent, and ingest your first PDF.
- Work through the tutorial. Build a table, check its sources, submit a canvas request, and review an agent's suggestion.
Open the Quickstart Learn the canvas workflow
Choose the right workflow¶
| You want to... | Read |
|---|---|
| View and extract a PDF without an ANCHOR API key | Quickstart |
| Keep ANCHOR model processing local, or configure an endpoint | Provider setup |
| Use a working folder or keep several projects | Environments and projects |
| Upload documents, create canvases, and inspect row sources | Documents and canvases |
| Connect Codex, Claude, Cursor, or another MCP client | Agent setup and manual configuration |
| Understand Verified, Unverified, and Stale rows | Claim and evidence |
| Script operations or discover agent tools | CLI reference and MCP reference |
| Understand storage and service boundaries | Projects, architecture, and on-disk substrate |
What to expect¶
PDFs are stored as originals plus derived page extraction and, when produced, interpreted regions. Local Docling extraction works without a model endpoint. Harness ingestion lets your agent interpret pages. Endpoint-backed ingestion uses your selected model service. Gold-region semantic search needs gold regions and embeddings.
Source links make an answer inspectable. Their precision varies from a whole page to a cell or explicit box. A citation or an automated evidence match does not certify that a value is suitable for an engineering decision.
FMU simulation needs the optional runtime. CAD and SysML tools are experimental. The default HTTP server is unauthenticated and binds to loopback; keep it local unless you provide an authenticated deployment layer.