Almost every mutation here comes in a pair — a preview that shows the candidate and the old and new values, then an apply you have to ask for — which is what makes handing an agent a personal work record tolerable. The payoff is on the read side: a task next month can find what was decided this month instead of starting the argument again.
A thin control plane over Granoflow, a local-first app for planning work, reviewing finished tasks and turning durable lessons into review cards. The server resolves the app's Local HTTP API and forwards structured requests to it; the app owns the data. It is not a code analyser or a CI fixer — it is the layer that records what task an agent is on, what happened, and what should be remembered.
- Tasks created from a requirement as it is discussed, listed, updated, completed, exported, validated and re-imported, with attachments and structured work nodes underneath them
- Projects and milestones created and resolved, with their descriptions kept as a living context map — the active milestone as the current phase, archived ones as final snapshots
- Completion records and review cards written at the end of a task, so the lesson survives the session it was learned in
- Durable experience as its own layer: experiences listed, searched, merged, linked to the work that used them, and unlinked with the impact shown first
- Knowledge assessed and materialised through a preview-then-apply pair at every step, with a per-task knowledge pack, its references, and an audit of what was actually adopted
- Evidence stored, searched and edited alongside it, so a claim in a review points at something
- A context pack for the current work and candidates from historical tasks, which is what lets a later agent continue rather than restart
- Bundled workflow skills read straight from the package — daily review, first-run import, task authoring, milestone coordination, portfolio orchestration and more — each returning its own SKILL.md plus a reference manifest, read-only, with no API token needed
- Setup that diagnoses itself: connection status, local API detection, a config write previewed before it happens, and an explicit report when an environment variable is shadowing the saved value instead of pretending it is active
- An escape hatch to the Local HTTP API for anything the typed tools do not cover
Node.js 20 or newer and the Granoflow desktop app running with its Local HTTP API enabled — this talks to your machine, not to a hosted service. The API defaults to http://127.0.0.1:56789, overridable with GRANOFLOW_API_BASE_URL, and the token comes from the client environment as GRANOFLOW_API_TOKEN. Non-secret connection defaults live in ~/.config/granoflow-mcp/config.json, with GRANOFLOW_MCP_CONFIG_PATH pointing elsewhere; tokens are never written there. Granoflow's local features are free, and without a membership the data stays on the device.
One command plus a key — npm install -g @granoflow/mcp-server, then supply credentials