ANCHOR

Agent-Native Canvas to Help Organize Resources
Source-Grounded Knowledge Canvas for Traceable Engineering Document Extraction

ANCHOR is a tool that lets you and your agent work with engineering documents.

Drop a PDF onto a canvas. The agent reads it and pulls the values you need into a spec table. Every value links back to the page and bounding box it came from, so you can click and see the source.

Drop FMU simulation models onto the same canvas and wire the extracted values into their parameters.

It runs on your laptop. A project is a folder: run anchor init in it, and its documents and canvases live in a hidden .anchor_data/ right there. Agents talk to it over MCP, so it works with Claude Code, Cursor, Claude Desktop, or any MCP client. There's an HTTP API and a CLI too.

First five minutes: docs/getting-started/tutorial.md.

---

Install

Two paths, depending on whether you want to use ANCHOR or hack on it.

Use it (from PyPI)

uv tool install anchor-kb

anchor and anchor-mcp are now on your PATH globally. The wheel includes the prebuilt frontend, so no Node toolchain is required to just run it.

The web UI ships prebuilt in the PyPI wheel. If you want the latest main and PyPI is behind, install from a local checkout. Installing directly from git source (uv tool install "git+https://github.com/Novia-RDI-Seafaring/anchor@main") does not build the frontend and fails at the wheel build hook. Build the frontend first:

git clone https://github.com/Novia-RDI-Seafaring/anchor && cd anchor
pnpm --dir web install --frozen-lockfile
pnpm --dir web build
uv tool install --force .

See Install from source for pnpm-not-on-PATH fallbacks.

If you want LLM-backed gold region extraction on your first PDF upload, create a .env file before starting ANCHOR; see Enable gold region extraction. Installation itself does not require an API key.

anchor serve              # -> http://127.0.0.1:8002

Requires Python >= 3.12. CI tests Linux and runs CLI smoke checks on macOS and Windows; verify browser and PDF workflows on your target platform.

If you prefer plain pip:

pipx install anchor-kb
# or, in a virtualenv:
pip install anchor-kb

Optional extras:

Extra	Install	Adds
fmus	uv tool install 'anchor-kb[fmus]'	FMU simulation runtime (fmpy). Without it, FMU tools fail closed unless you opt into the synthetic demo with ANCHOR_FMU_DEMO=1.

Hack on it (from source)

git clone https://github.com/Novia-RDI-Seafaring/anchor
cd anchor
uv sync --extra dev          # adds pytest, ruff, import-linter
pnpm --dir web install

Start the backend in one terminal:

uv run anchor serve

Start the frontend development server in a second terminal:

pnpm --dir web dev

Open http://localhost:5173 for the development UI. The backend remains on http://127.0.0.1:8002.

Source development requires Node.js 20+ and pnpm 10. If pnpm is not installed globally, use the Corepack form instead: corepack pnpm@10 --dir web install and corepack pnpm@10 --dir web dev. If Corepack is blocked but Node.js and npm are available, use npx pnpm@10 --dir web install and npx pnpm@10 --dir web dev.

For normal use, run anchor init in a project folder. It writes an anchor.toml marker and a hidden .anchor_data/ there, and binds the project to an environment (the provider and data zone). Commands run inside the folder then resolve it automatically. Configuration precedence is: explicit flags, ANCHOR_* environment variables, the project anchor.toml, the environment env.toml, then built-in defaults. Provider, endpoint, and local-only policy are environment-owned security settings: projects cannot redirect or weaken them, and process variables cannot retarget a named environment.

Releases are tag-driven: pushing a v* tag triggers the release workflow, which publishes to PyPI via OIDC trusted publishing (no token sits in the repo). See PUBLISHING.md for the full release process.

---

Quick start

Nothing to a source-grounded value in about five minutes, no API key:

uv tool install anchor-kb
anchor env create home --provider harness --yes   # no Anchor-side model endpoint
anchor install claude-desktop --env home          # or claude-code / cursor
# restart your harness, drag a PDF into the chat, ask it to ingest
anchor serve                                       # http://127.0.0.1:8002

On the harness provider your agent reads the pages, so you get gold regions (values with page + bbox provenance) with no Anchor API key. Page content is visible to the connected harness and may reach its model provider. Use only a harness whose data policy is approved for the document. Full walkthrough, provider decision table, and troubleshooting: Quickstart.

Then docs/getting-started/tutorial.md walks the anchor demo -> "agent fills the placeholders" tour.

Drag-drop, CLI ingest, and harness ingest

There are two PDF ingestion modes:

Mode	How it starts	Who extracts gold regions	When to use it
Built-in ingest	Canvas drag-drop, HTTP upload, MCP ingest_pdf, or CLI anchor ingest	ANCHOR's configured extractor	Normal uploads, scripted ingest, and projects with a configured vision endpoint
Harness-driven ingest	MCP ingest_begin -> ingest_get_page -> ingest_submit_page -> ingest_finalize	The connected agent harness, page by page	Provider harness, no-key workflows, or difficult PDFs where extraction quality matters more than speed

Both modes publish to the same project data layout: bronze/, silver/, gold/, and canvas state under .anchor_data/. A document ingested without a vision endpoint may have silver data but no gold regions. Configure a provider and re-ingest with --force, or use harness-driven ingest.

That's the whole loop. Every PDF you ingest becomes a structured set of regions on disk; every canvas you create is a folder you can zip and email.

---

Using ANCHOR with an AI agent

ANCHOR exposes its tools over MCP (Model Context Protocol). For Claude Code, the quickest path is the plugin marketplace. It needs no prior install of the anchor CLI, only uv:

/plugin marketplace add Novia-RDI-Seafaring/anchor
/plugin install anchor@anchor

The plugin registers the MCP server (via uvx --from anchor-kb anchor-mcp) and the anchor skill in one step. See the Claude Code plugin guide for details.

Alternatively, with the CLI already installed, register the local stdio server with:

anchor install claude-code

Pick one of the two paths, not both.

Open Claude Code inside a folder configured with anchor init. In any conversation, run /mcp and you should see anchor listed with its available tools. The exact list depends on optional extensions such as FMU support. Then talk normally:

"Ingest the PDF at ~/Downloads/lkh-pump.pdf and create a canvas called pump-analysis with a document node for it."

"What does the document say about max inlet pressure for the LKH-5 at 50 Hz? Place the answer as a fact card on the pump-analysis canvas, with an evidence edge back to the source page."

Claude calls the MCP tools directly. Your browser tab on localhost:8002/c/pump-analysis, if open, sees nodes appear live via SSE. Multi-client real-time sync between agents and humans is the default.

For Cursor:

anchor install cursor

anchor install claude-code / cursor wire the MCP server for the default environment. For Claude Desktop, or to serve a specific environment, use anchor install claude-desktop --env <name>; it writes a named entry, echoes the data zone before wiring, and is collision-safe. See the agent setup guide.

See the agent configuration guide for Codex, OpenCode, Cursor, Claude Code, and generic stdio examples.

---

Where data lives

A project is a folder. Its corpus and canvases live in a hidden .anchor_data/ beside an anchor.toml marker. Everything is plain files: tar it, mail it, diff it in git.

your-project/
├── anchor.toml             # binds this folder to an environment (provider + data zone)
└── .anchor_data/
    ├── bronze/<original>.pdf   # raw PDFs, flat (original filename)
    ├── silver/<slug>/          # Docling extraction + per-page markdown + page PNGs
    ├── gold/<slug>/            # structured regions with page + bbox provenance
    └── canvases/<slug>/        # meta.json, state.json, events.jsonl (append-only log)

This layout is the contract. You can hand-edit the JSON, copy a canvas folder to another machine, or version-control the whole project. The file-level detail is in On-disk substrate.

A project created by an agent (no working folder) is managed under its environment at ~/.anchor/envs/<env>/projects/<name>/, with the same .anchor_data/ inside. A pre-existing ~/anchor-data from older versions keeps working until you run anchor migrate.

---

Configuration

Provider, model, and data-zone settings live in an environment's env.toml, created with anchor env create <name>. Choose this environment before uploading a document. The provider is a data-handling decision: it determines whether page content stays on the computer, goes to a harness, or goes to a configured model endpoint.

Provider	Gold regions	Page destination	Key needed
local	No	This computer only	No
ollama	Yes	Your Ollama server, normally this computer or LAN	No
harness	Yes	The connected agent harness	No ANCHOR key
openai	Yes	OpenAI	Yes
azure	Yes	Your configured Azure OpenAI endpoint	Yes
custom	Yes	Your configured OpenAI-compatible endpoint	Yes

Create and verify one environment explicitly:

anchor env create private-gold --provider ollama --vision-model llava --yes
anchor check --env private-gold --probe

For all provider recipes and the security implications of each choice, read Choose a provider and enable gold before processing sensitive documents.

env.toml is not .env

For an environment named private-gold, the supported files are:

~/.anchor/envs/private-gold/env.toml   provider, endpoint, models, data zone
~/.anchor/envs/private-gold/.env       optional endpoint credential only

env.toml is non-secret and selects the provider. The environment's .env is gitignored and should contain only secrets such as:

ANCHOR_OPENAI_API_KEY=<credential-for-this-environment>

The key does not select a provider and does not authorize egress by itself. ANCHOR loads the environment .env only after that environment has a valid env.toml. If the provider is missing, local, or harness, the runtime does not build an endpoint client even when a key is present. Ollama does not need a user key. Azure and custom endpoints require ANCHOR_OPENAI_API_KEY.

Do not put OPENAI_API_KEY=... in the environment .env; its scoped loader imports only ANCHOR_ variables. A process-level OPENAI_API_KEY is accepted only for the public openai provider with no custom base URL. Using ANCHOR_OPENAI_API_KEY consistently is clearer and keeps the credential scoped to the named environment.

For openai, azure, and custom, omit --yes during interactive setup. ANCHOR asks for the key with hidden input and saves it to the selected environment's .env. With --yes, no prompt runs and no key is saved; create the environment .env separately. Never put a key directly in a command-line argument because shell history may retain it.

Run anchor init --env <name> to bind a working folder to an existing environment. It writes a non-secret project anchor.toml and a hidden .anchor_data/. A project cannot redirect its environment's provider or endpoint.

Enable gold and recover a silver-only document

Gold regions are created during ingestion only. After creating or changing an environment, restart the server so it resolves the new provider:

anchor serve --env private-gold --project default

If the server is already running, stop it with Ctrl+C first. Documents already ingested as silver-only are not backfilled automatically. Select the correct environment and reingest the original PDF:

anchor use private-gold default
anchor ingest "/path/to/document.pdf" --force
anchor list

anchor list should report "has_gold": true and a non-zero region_count. With a paid endpoint, --force repeats billable model calls. See the provider guide for Windows commands and a complete troubleshooting checklist.

Local-only / no-egress mode (confidential documents)

For confidential documents that must never leave the host, run the local provider. It runs Docling layout extraction and OCR with no external model calls: no OpenAI-compatible client is built for any stage, regardless of any key in your environment, and model loading is pinned offline. Gold region extraction is skipped. Semantic region embeddings are therefore also absent, because the current embedding pipeline operates on gold regions. Bronze, silver, page text, and rendered pages remain available.

# One-time: warm the local model cache while you still have network.
anchor models prefetch                  # downloads bge-small + docling models

# Create a no-egress environment and bind a project to it.
anchor env create vault --yes --provider local
cd confidential-project && anchor init --env vault

# Verify the posture before feeding sensitive input.
anchor check --env vault                # shows "local-only: ON - no external egress"

# Ingest with no outbound connections.
anchor ingest "C:\path\to\datasheet.pdf"

The local provider records local_only = true in the environment's env.toml, which the runtime honors identically across the CLI, HTTP and MCP adapters (so an agent-launched anchor-mcp gets the same no-egress posture). For a confidential corpus, use Anchor's built-in local ingest path. Do not hand the document to a cloud-backed harness or ask it to perform harness-driven ingest, because the harness itself would then receive the page content. On a fully locked-down host, also export the HuggingFace offline switches so a cache miss fails fast instead of attempting a download:

export HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1

anchor models list reports the exact model set a local-only ingest needs; anchor check echoes it and whether the offline env is active.

---

Commands

Run most commands from inside a project folder and they resolve it automatically (via the anchor.toml marker). Full reference: docs/reference/cli.md.

# Environments (the provider / data-zone profile = the trust boundary)
anchor env create NAME [--provider local|harness|ollama|openai|azure|custom] [--base-url ...] [--vision-model ...]
anchor env list | show NAME | default NAME | set-description NAME "..."

# Projects (a folder = a corpus + its canvases)
anchor init [NAME] [--env NAME] [--provider ...]  # start a project in this folder
anchor project create NAME [--env NAME]          # a managed project under the env
anchor project list | set-description | move NAME --to ENV
anchor use ENV [PROJECT]                          # session default so you can omit --env
anchor migrate                                    # fold a legacy ~/anchor-data in
anchor check [--env NAME] [--probe] [--fix]       # audit the data zone before ingesting

# Local models (offline / no-egress provisioning)
anchor models list [--env NAME]                   # the local model set an ingest needs
anchor models prefetch [--env NAME] [--embed-model M]  # cache them while you have network

# Documents + search
anchor ingest PDF_PATH [--skip-polish] [--skip-regions] [--force]
anchor list | index SLUG | regions SLUG [--page N] | page-text SLUG PAGE
anchor embed [SLUG] [--overwrite] | search "<query>" [--k N]

# Canvas + server
anchor serve [--env NAME] [--project NAME] [--host HOST] [--port PORT]
anchor demo  [--no-serve]
anchor canvas list | create SLUG [--title TITLE] | placeholders SLUG | snapshot SLUG

# Agents (write a named MCP pointer for an environment)
anchor install claude-desktop --env NAME [--name ENTRY] [--create]
anchor install claude-code [--env NAME] | cursor [--env NAME] | print

# Extensions (OIP producers) + misc
anchor extensions list | info NAME | add MANIFEST | remove NAME | discover | schema
anchor version

anchor-mcp --env NAME runs the MCP server over stdio for one environment (used by an agent's MCP harness; you don't normally invoke it yourself). --data-dir is still accepted on the document/canvas commands to point at a raw storage dir, but the project folder is the usual way in.

---

Architecture (one paragraph)

ANCHOR is a hexagonal modular monolith. Pure domain code in core/ (no I/O, no framework imports - enforced by lint-imports). Concrete protocol implementations in infra/. Transport adapters in adapters/ (HTTP, MCP, CLI, SSE). The Python wheel ships the React frontend bundle inside it (anchor/_web_dist/) so one process serves both the API and the UI. State changes are events, persisted to events.jsonl per canvas, broadcast to subscribers (agents on MCP, browsers on SSE). See the architecture docs.

---

Extensions and the Open Ingestion Protocol

ANCHOR's canvas is one OIP consumer. PDF ingestion is one OIP producer, bundled with this build. The protocol, specified at github.com/Novia-RDI-Seafaring/OIP, is governance-neutral: any tool that produces ingested knowledge in OIP shape can plug in, and any OIP-aware consumer can read its output. A transcription tool, a code-region extractor, a web crawler, or your own ingestion logic does not need to import ANCHOR. It only needs to ship an OIP manifest at a known location.

The CLI surfaces this:

anchor extensions list                        # what producers can this ANCHOR see?
anchor extensions discover                    # where does it look for manifests?
anchor extensions add <path-to-manifest.json> # register a new producer (system-wide)
anchor extensions schema                      # print a starter manifest to edit
anchor extensions info anchor-pdfs            # full manifest for one producer

Discovery, in priority order:

1. Per-data-dir: <data-dir>/.oip/producers.d/*.json (highest priority; bound to a specific workspace tree)
2. System-wide: ~/.config/oip/producers.d/*.json (any installer can drop a manifest here; visible to every OIP consumer on the machine)
3. Bundled: compiled into this ANCHOR wheel (anchor-pdfs, anchor-fmus, and anchor-cad; SysML tools are also exposed by the bundled MCP server)

For implementation status: today, an OIP-registered producer is visible in extensions list but ANCHOR doesn't yet spawn external producer MCP servers and proxy their tools. That's the next engineering lift. See the OIP repo for the spec and EXTENSIONS.md for ANCHOR's host-side roadmap.

---

Tests

uv sync --extra dev                       # one-time: install pytest/ruff/import-linter
uv run pytest                             # ~570 backend tests
uv run lint-imports                       # 6 dependency-rule contracts
pnpm --dir web test                       # ~180 web tests (Vitest)
pnpm --dir web exec tsc --noEmit          # web typecheck

The test seam is function-based pytest with in-memory implementations of every port. Real I/O tests use tmp_path. The frontend tests cover canvas primitives, the SSE event store, and the inline-edit hooks.

---

Status & roadmap

v0.2 (current): canvas primitive + PDF ingestion in one package, real-time SSE sync, MCP integration, folder-based projects under named environments (the data-zone / trust boundary), skill + pointer installers for Claude Code / Cursor / Claude Desktop, backend and web test suites, hexagonal contracts enforced.

Near-term: complete remaining node renderer and asset workflows, then stabilise the extension registration surface.

Mid-term: split the canvas primitive (anchor-canvas) and PDF extension (anchor-canvas-pdfs) into separately-publishable packages, and stabilise the extension contract for third-party authors.

Longer term: other ingestion extensions (audio/video transcription, code, web), shared org docs / personal canvases topology, optional Postgres event store for very large workspaces.

---

Security model: read before exposing

ANCHOR's HTTP server is unauthenticated by design. It edits local engineering data (workspaces, documents, FMU files) and is meant to run on your own machine.

• Default bind is 127.0.0.1 (loopback). Nothing else on the LAN can reach it unless you pass --host 0.0.0.0.
• CORS is restricted to the dev Vite origin (localhost:5173); set ANCHOR_CORS_ORIGINS=https://your-host for explicit overrides.
• Workspace slugs and upload filenames are policy-checked and containment-asserted before they hit disk. The v2 codebase does not trust client-supplied paths.

If you want to share an ANCHOR instance on a network, add your own reverse proxy with auth in front of it (Tailscale, OAuth proxy, basic-auth nginx, ...). Don't expose the unauthenticated port directly.

Limitations (v0.2)

These extensions are intentionally rough; we ship them so you can see the shape and contribute, not as finished features:

• anchor_cad: parametric-CAD producer (jscad/openSCAD) ships as a proof of concept; full feature parity with STEP/STL viewing is on the roadmap. SVG export still has a known font-handling bug.
• anchor_sysml: SysML import (BSD-3-Clause fixtures from the OMG reference) and export to SVG/markdown are experimental; we'll swap the hand-rolled IR for the official Pydantic model when that lands.
• anchor_fmus: FMU simulation requires fmpy (install via uv tool install 'anchor-kb[fmus]'). Without it the extension fails closed; set ANCHOR_FMU_DEMO=1 to use the synthetic-output runtime (every result is stamped synthetic=true so the UI can warn you).

License

MIT, see LICENSE.

Citation

If you use ANCHOR, please cite the software repository:

@misc{ANCHOR,
  author       = {Lamin Jatta and Christoffer Bj{\"o}rkskog and Mikael Manng{\aa}rd and Johan West{\"o}},
  title        = {ANCHOR: Agent-Native Canvas to Help Organize Resources for Traceable Engineering Document Extraction},
  year         = {2026},
  howpublished = {\url{https://github.com/Novia-RDI-Seafaring/anchor}},
}

GitHub-compatible citation metadata is provided in CITATION.cff.

Acknowledgments

This work was done in the Business Finland funded project Virtual Sea Trial.

Contributing

Open changes as short-lived branches targeting main; see CONTRIBUTING.md. Run uv run --extra dev pytest and uv run --extra dev lint-imports before pushing backend changes. See EXTENSIONS.md for the proposed third-party extension contract and its current implementation status.