A thin front-end to your local-first granary cluster, alongside the CLI, GUI, and TUI.
Ask the MoE router, query the knowledge brain, and see the board — all against local
models (Ollama / llama.cpp / LiteLLM). No cloud.
Commands (Cmd/Ctrl-Shift-P)
granary: Ask the cluster (infer) — routes your prompt through the MoE (granary infer).
granary: Ask about the selection — right-click a code selection → ask about it.
granary: Query the knowledge brain — grounded, cited retrieval (granary knowledge query), abstains when unsure.
granary: Ask the silo — search your local data brain (granary grainsilo ask): a grounded answer, or an honest "no confident match" instead of a guess.
granary: Show the board / List persona agents / Cluster status.
granary: Register as an MCP server — writes .vscode/mcp.json so MCP-aware chat/agent
extensions get granary's tools (knowledge_search, knowledge_query, agents_list, agents_ask).
Two ways it talks to granary
CLI (default) — the commands above shell out to the granary CLI. Works everywhere.
MCP — granary already speaks MCP (granary mcp serve, stdio). Run granary: Register as
an MCP server and any MCP host (VS Code's built-in MCP, Copilot Chat, Claude Dev, etc.) can
call granary's tools directly. The generated .vscode/mcp.json:
granary.command — how to invoke granary. Default granary. Set to an absolute venv path
(/…/granary-harness/.venv/bin/granary) or python -m granary if granary isn't on PATH.
granary.ground — ground Ask answers in the indexed repo (RAG).
granary.knowledgeNamespace — optional namespace(s) for knowledge queries.
Run it (dev)
code editors/vscode (open this folder).
Press F5 → an Extension Development Host opens with the extension loaded.
Make sure granary runs in a terminal first (granary status); set granary.command if needed.
Package (optional)
npm i -g @vscode/vsce && vsce package → a .vsix you can install with
Extensions: Install from VSIX. (Plain JS — no build/bundle step.)
