For the complete documentation index, see llms.txt. This page is also available as Markdown.

MCP server

Connect AI assistants and agents to Smabbler through the Model Context Protocol.

Smabbler has an MCP server, so AI assistants and agents that support the Model Context Protocol can work with your graph models directly: upload and analyze text, build and switch on models, retrieve knowledge and ask Smabbler's agent, all with your own account.

Server address

https://beta.mcp.smabbler.com/mcp

The server uses the Streamable HTTP transport. This is the beta server; it works with your beta Portal account.

Signing in

You sign in with your Smabbler Portal account, the same e-mail and password as in the Portal. You do not need an API key. When you connect a client, it opens the Smabbler sign-in page in your browser; if you are already signed in to the Portal there, you are connected at once.

Everything the client does happens in your account: sources and models created through MCP appear on the Files and Models pages, and GCUs are used from your balance.

Connecting a client

  • Claude (web and desktop app): open Settings → Connectors, choose Add custom connector, paste the server address and click Connect, then sign in.

  • Claude Code: add the server with the command below, then run /mcp in Claude Code and choose the server to sign in.

  • Other MCP clients: add a remote MCP server with the address above. The client must support sign-in with OAuth; the server publishes everything the client needs to register itself.

claude mcp add --transport http smabbler https://beta.mcp.smabbler.com/mcp

Tools

The server offers 16 tools. You do not call them yourself: describe what you want, and the assistant picks the tools.

Knowledge

  • smabbler_retrieve_knowledge - searches one or more active models and returns the ranked results, each with the source row it comes from. The assistant writes the answer from them. This is the main way to get knowledge out of a model.

  • smabbler_ask_agent - lets Smabbler's agent answer the question in Markdown, optionally with sub-queries, planning and synthesis hints and the earlier questions and answers of the conversation.

  • smabbler_wait_for_result - keeps waiting for a search or an answer that took longer than about a minute.

Sources

  • smabbler_upload_source - uploads short text from the conversation as a TXT, MD or CSV source, up to about 50 KB, optionally with tags. Upload files on the Files page instead; the assistant can then find them.

  • smabbler_analyze_source - analyzes a source: TXT and MD as text, CSV with an ID column and text columns, optionally with the column names. Analysis uses GCUs.

  • smabbler_list_sources, smabbler_get_source - list the sources and show one source with its status and efficiency.

  • smabbler_update_source_tags - replaces the tags of a source.

  • smabbler_delete_source - deletes a source. This cannot be undone.

Models

  • smabbler_build_model - builds a new model from analyzed sources.

  • smabbler_set_model_active - switches a model on or off. Only an active model can be searched.

  • smabbler_edit_model_sources - adds or removes sources of an existing model.

  • smabbler_list_models, smabbler_get_model - list the models and show one model with its status, sources and history.

  • smabbler_delete_model - deletes a model. This cannot be undone.

Account

  • smabbler_get_gcu_balance - shows your remaining GCUs.

A typical session

  1. Upload a file on the Files page, or ask the assistant to upload a short text.

  2. Ask the assistant to analyze the source and build a model from it. Analysis takes about a minute per source.

  3. Ask it to switch the model on. Activation takes a few minutes.

  4. Ask your questions. The assistant searches the model and answers from what it finds.

  5. When you are done, ask the assistant to switch the model off.

Good to know

  • A question or search text can be up to 500 characters.

  • A search or an answer waits up to about a minute. If it takes longer, the assistant gets a handle and waits for the result with smabbler_wait_for_result.

  • A model that was not used for 30 minutes goes to sleep. The next search wakes it up, and the first answer can take a few minutes.

  • Deleting a source or a model cannot be undone.

  • Without an MCP client, use the Smabbler API with an API key.