Skip to content

Projects

A project is a folder with an anchor.toml in it. That one file makes the folder self-describing: it records the data directory, the AI provider (and so the data zone), and the models. Every ANCHOR adapter — the anchor CLI, anchor serve, and an agent-launched anchor-mcp — resolves the project from that file, so you configure once and run from inside the folder.

Create one with anchor init

cd ~/my-project
anchor init

anchor init asks the question that matters first — where may document content go? — and writes a non-secret anchor.toml. The choice of provider is the choice of data zone:

Provider Data zone
local on-host; nothing leaves the network (bronze/silver + local search; no gold regions)
ollama your machine / LAN; no internet egress, with offline gold regions via a local vision model
openai public cloud
azure your Azure tenant / region
custom any OpenAI-compatible endpoint; you label the zone

It then picks the embedding model and the data directory (default <project>/anchor-data), and prints the next steps. The API key is never written to anchor.toml — keep it in ANCHOR_OPENAI_API_KEY or a gitignored .env so a committed config carries no secret.

How adapters find the project

anchor.toml is discovered by walking up from the working directory, or by an explicit ANCHOR_CONFIG path. So:

  • CLI / server — run anchor ingest, anchor serve, etc. from inside the folder (or a subfolder) and they use the project automatically. No repeated --data-dir.
  • MCP / agents — anchor-mcp resolves the project from the directory the agent launches it in. Name one explicitly with anchor-mcp --project <folder>.

Because resolution happens at runtime, you register the MCP server once (anchor install claude-code, with no baked data dir) and it serves every project: open the agent in a project folder and its tools target that project. No reinstall when you switch projects.

{ "mcpServers": { "anchor": {
    "type": "stdio", "command": "anchor-mcp"
} } }

Configuration precedence

Highest priority first:

  1. Explicit flags / constructor args (e.g. --data-dir)
  2. ANCHOR_* environment variables
  3. A .env file
  4. The project anchor.toml
  5. Built-in defaults (data dir ~/anchor-data, embeddings bge-small, …)

So an operator's ANCHOR_* override always wins over a committed project default, and a malformed anchor.toml is ignored with a warning rather than crashing the CLI.

Data zones and egress

The provider you pick governs what leaves the host:

  • local / ollama keep document content on your machine or LAN.
  • openai sends page images and extracted text to OpenAI.
  • azure / custom send the same content only to the endpoint you name.

Embeddings stay local (bge-small) by default, so text never leaves the host even when the vision model is remote; choosing a text-embedding-* model opts those vectors into the endpoint.

See Configuration for the full key reference and Agent setup for connecting a harness.