Comparison is the point: focus puts one question to several models across rounds, and a workflow runs steps over different providers and hands the outputs to a judge with a table beside them. Two smaller things are worth knowing — a local model can take part through Ollama or LM Studio at no cost, and preview_prompt_technique shows what a technique would do to your prompt before a call is spent on it. Which tools you actually see depends on which keys you set.
A multi-model orchestration server. One entry point routes a request to whichever provider suits it, a focus mode runs several models against the same question over multiple rounds, and workflows chain steps across providers with a comparison table and an optional judge at the end.
- tachi takes a plain-language request and routes it to the right mode; focus runs multi-model reasoning over one question with the models, rounds and style you pick
- nextThought does sequential thinking — each thought logged, optionally executed by a named model, with context distillation between steps and a final judge at the end
- local_query sends a prompt to a local open-weight model through Ollama, LM Studio, llama.cpp or vLLM: offline, private and free, configured with LOCAL_LLM_BASE_URL and LOCAL_LLM_MODEL
- Workflows created, listed, visualised and executed, with workflow_start, continue_workflow and workflow_status turning a run into a resumable session
- validate_workflow and validate_workflow_file check a workflow's YAML or JSON before it runs, including whether its ${step.output} interpolation references point at anything real
- list_prompt_techniques, preview_prompt_technique and execute_prompt_technique let you see how a prompting technique rewrites your prompt before spending a call on it — preview returns an execution token you can run afterwards
- doctor reports which API keys were detected, which tools are visible and which are hidden and why, and which profile is active
- usage_stats shows or resets per-tool usage counts
- TACHIBOT_PROFILE chooses how much of the surface loads: minimal, research_power, code_focus, balanced or full
Published as tachibot-mcp and launched over stdio, on Node.js 22 or newer. The provider tools are key-gated: PERPLEXITY_API_KEY for search and research, GROK_API_KEY for xAI, and OPENAI_API_KEY are listed as required, with GOOGLE_API_KEY and OPENROUTER_API_KEY optional. TACHIBOT_PROFILE selects the loaded tool set and defaults to balanced. A tool whose key is absent hides itself, and doctor is what tells you which ones and why.
One command plus a key — npm install -g tachibot-mcp, then supply credentials