Configuration

Configuration is loaded and validated by src/config.ts using a Zod schema. Values are read from environment variables, with .env files loaded in the following order (later values override earlier ones):

1. .env in the current working directory (if present).
2. .env in the project root (fallback).

Required variables

Variable	Description
ANTHROPIC_API_KEY	Required. Anthropic is the only supported LLM provider (src/utils/model.ts); there is no fallback to another provider.
PATBA_API_KEY (or API_KEY / AWS_APP_API_KEY)	Shared secret used to authenticate REST API requests (see the X-API-Key header in Server Usage).

validateConfig() throws at startup if no LLM key or no PATBA_API_KEY is configured, for every entrypoint (index.ts, cli.ts, server.ts, mcp.ts).

Optional variables

Variable	Default	Description
ANTHROPIC_MODEL	claude-sonnet-5	Claude model name used by the agent.
ANTHROPIC_MAX_TOKENS	16000	Output tokens per reply. Must be stated: the client derives its own default from a table of model-name prefixes that claude-sonnet-5 is not in, and falls through to 4096. Raising it far past the default does not work either — the Anthropic SDK refuses a non-streaming request estimated to run over ten minutes, which caps this around 21333.
NODE_ENV	development	Runtime environment name.
SQLITE_DB_PATH	state.db	Path to the local SQLite checkpoint database (relative paths are resolved against the project root).
PORT	8080	Port used by the Express REST/SSE server.
AWS_REGION	sa-east-1	AWS region used by the S3 storage client.
S3_BUCKET_NAME (or AWS_S3_BUCKET)	pinky-and-the-brain-agents-state-store-dev-094094788286	S3 bucket used to persist state in cloud environments.
SLACK_BOT_TOKEN	unset	Enables the /webhooks/slack reply flow. Logged as an informational message (not a hard error) when missing.
LANGCHAIN_TRACING_V2 / LANGSMITH_TRACING	false	Enables LangSmith tracing.
LANGCHAIN_API_KEY / LANGSMITH_API_KEY	unset	Required for LangSmith tracing to actually send traces.
LANGCHAIN_PROJECT / LANGSMITH_PROJECT	pinky-and-the-brain-agents	LangSmith project name.
BLOG_REPO_PATH	../thiagocolen.github.io	Checkout of the blog whose articles MCP server publishes finished articles.
GEMINI_API_KEY (or GOOGLE_API_KEY)	unset	Enables cover-image generation, and the vector half of retrieval. Without it articles still publish (without an image) and retrieval still works (BM25 alone).
RETRIEVAL_MODE	hybrid	lexical uses BM25 alone. Hybrid already falls back on its own when there is no key, no embeddings file or a failed call, so set this only to drop the per-turn embedding call deliberately.
GEMINI_EMBEDDING_MODEL	gemini-embedding-001	Embeds chunks during npm run ingest, and queries at run time.
GEMINI_EMBEDDING_DIM	256	Dimensions kept from the model's native 3072. Changing it invalidates every stored vector — the agent notices the mismatch and ignores them rather than comparing different lengths, so re-run npm run ingest.
CURATED_CONTENT_PATH	sibling patb-rag-curated-content	Folder npm run ingest reads source documents from, one directory per area.
GEMINI_IMAGE_MODEL	gemini-3.1-flash-lite-image	Model used for cover images and article figures.
GEMINI_IMAGE_SIZE	1K	Pixel tier requested for every generated image. Which tiers a model serves varies — the default one accepts 1K and refuses both 512 and 2K — and a refused size is retried once at 1K, so a wrong value costs latency rather than pictures.

Publishing articles to the blog

The blog is a separate repository, and it owns its own publishing tools. When Pinky asks The Brain to publish a finished article, the agent does not write posts itself — it drives the articles MCP server that ships inside the blog at mcp-server/index.js, calling create_draft → add_asset → update_post → stage_changes.

This matters because it removes a whole class of drift. The blog's frontmatter shape, its slug rules, where images live and which branch is safe to push are now defined in exactly one place — develop-tools/posts.js, the same module the site's own npm scripts use — and this agent is a client of it rather than a second opinion about it.

Prerequisites

BLOG_REPO_PATH must point at a checkout that:

• is on a branch containing mcp-server/ (it arrives with feature/mcp-publisher-tool), and
• has had npm install run inside mcp-server/ once.

If either is missing, publishing fails with a message naming the exact path it looked for.

What publishing does and does not do

• The server writes to its own git worktree, .mcp-worktree/, checked out to new-articles. You can be mid-edit on any branch of the blog without colliding with it.
• Pushing uses that checkout's own git credentials. No token belongs in .env.
• Nothing is ever published live. Two independent gates stand between an article and the public site, and the agent passes neither on its own: the post's status flag (the agent never calls publish_post, so drafts stay unpublished) and the pull request itself. No workflow builds new-articles, so pushing there cannot reach the web.

Cover images are generated with gemini-3.1-flash-lite-image and added through add_asset, which returns the /images/… URL that goes into the post's cover_image. This is the only place a non-Anthropic model is used; all text generation remains Anthropic-only.

AWS credentials and the S3 fallback

src/storage/s3.ts lets the AWS SDK's default credential provider chain resolve credentials — environment variables, shared config, SSO/AWS_PROFILE, or a container/instance role. If nothing resolves, the service transparently falls back to an in-memory mock store, so local development works without any AWS credentials configured. The deployed Lightsail container service is not currently given a runtime IAM role (see Infrastructure), so this fallback is also what runs in production today.

API key masking

The REST server validates the X-API-Key request header against PATBA_API_KEY using validateApiKey() in src/config.ts; unauthorized requests receive 401 responses before reaching any route handler.