Bike4Mind

What it does

Bike4Mind is a multi-model AI workspace built around notebooks, a document knowledge base and data lakes, offered as a hosted service at app.bike4mind.com or as a self-hosted open-core stack. Its command-line package, @bike4mind/cli, is mainly an interactive agent terminal, but its `b4m mcp serve` subcommand turns it into an MCP server that talks to a Bike4Mind backend on the user's behalf.

In the published 2.0.0 release the server registers ten tools (checked in the package's bundled serve module): listing, reading and creating notebooks, sending a chat message into a notebook and waiting for the reply with its cited sources, semantic search over notebook content, listing data lakes, searching files and fetching a file's metadata with a signed download link, and generating a sound effect or speech audio. A lake-grounded answer is built by chaining list_lakes, create_notebook with a dataLakeId, then send_message. It also serves notebooks, files, projects and artifacts as readable resources.

List-style tools return 25 items by default (search_knowledge_base returns 10), up to 100 with `limit`. It runs over stdio by default, or as a stateless streamable HTTP endpoint with `--http --port <n>`.

What it needs

Node.js 24 or newer (the package's engines field). A Bike4Mind account, either on the hosted service or on a self-hosted instance, plus an API key created under Settings > API Keys with scopes for the tools you use (keys start `b4m_live_`); running `b4m login` once instead stores a sign-in the server can use. Audio generation spends the account's generation credits.

Limits

Version 2.0.0, the one on npm, has 10 tools. The repository docs list 17: rename_notebook, clone_notebook, delete_notebook, list_projects, get_project, create_project and generate_image, plus Briefcase prompts, arrive in 2.1.0, which its changelog records but npm does not yet carry. Resource listing is capped at 100 entries per template with no paging, and a failed listing comes back as an empty list with the error on stderr. HTTP mode binds only 127.0.0.1 and serves only /mcp; there is no flag to change the bind address. generate_sound_effect accepts 1-1000 characters of text and 0.5-30 seconds of length.

Who it suits

Suits people who already keep notes, documents or data lakes in Bike4Mind and want an assistant in another client to search them, start notebooks and get cited replies from Bike4Mind's own models. It is of no use without a Bike4Mind account or self-hosted instance. Because tools can create notebooks and spend credits on chat and audio generation, a key with only read scopes suits read-only use, though the docs note most routes do not enforce scopes. People who want the rename, delete, project or image tools should wait for a release after 2.0.0.