Configuration

AIProfile uses these defaults when no config.yaml is present. Create or edit config.yaml to change settings:

server:
  port: 3000 # HTTP port the MCP server listens on

owner:
  name: null # learned through ingestion
  preferred_language: null

llm:
  provider: node-llama-cpp
  model_path: ./models/qwen3-4b-instruct-q4_k_m.gguf
  temperature: 0.2
  max_tokens: 1200
  context_tokens: 4096

memory:
  path: ./memory
  mode: encrypted # default; set to plain only for local testing/debugging

auth:
  mode: local # default; local OAuth grants backed by encrypted storage
  anonymous_enabled: true # unauthenticated clients can only use public-safe ask
  # issuer: https://abc123.ngrok-free.app
  # resource: https://abc123.ngrok-free.app/mcp

safety:
  allow_first_person: true
  public_can_access_private_memory: false
  require_disclaimer_for_inferred_answers: true

Server

server.port controls the local HTTP port. The MCP endpoint is:

http://localhost:<port>/mcp

Local model

llm.model_path points to the GGUF model file. Use Model Setup to download a curated model and optionally update config.yaml.

llm.max_tokens controls the response token budget for local model calls.

llm.context_tokens is the model context window used for prompt budgeting. Ingest rejects oversized source text before calling the model when the estimated prompt would exceed this window. If you use a model with a larger context window, increase this value to match the model.

Memory mode

memory.mode: encrypted is the default and recommended setting. It stores memory in an encrypted SQLite database.

memory.mode: plain stores memory as an unencrypted on-disk SQLite database. Use plain mode only for local testing and debugging.

Authentication

auth.mode: local enables local OAuth grants and access tokens backed by the encrypted memory vault. In this mode, encrypted memory is required.

auth.mode: off disables token checks and exposes the full MCP tool surface without authentication. Use it only for isolated local testing.

auth.anonymous_enabled: true documents the intended anonymous access behavior. Anonymous access is limited to public-safe ask.

auth.issuer is the externally visible OAuth issuer origin. Set it to the public HTTPS tunnel origin for ChatGPT or other web-hosted clients.

auth.resource is the externally visible MCP URL. Access tokens are audience-bound to the resource URL, so update this whenever your ngrok URL changes.

Non-interactive startup

For service managers or MCP clients that start the server without an interactive terminal, provide the memory password through an environment variable:

AIPROFILE_PASSWORD='your memory password' npm start
AIPROFILE_PASSWORD='your memory password' npx --yes --ignore-scripts=false aiprofile serve

Or store the password in a local file and pass the file path:

npm start -- --password-file ./local-password-file
npx --yes --ignore-scripts=false aiprofile serve --password-file ./local-password-file

Keep password files outside version control.