io.github.pathmodeio/mcp-server

Run the deterministic preflight on an intent spec before handing it to an implementation agent. Six calibrated checks — title, objective, outcomes, constraints, edge cases, verification — computed by pure functions: no model call, no network, the same spec always gets the same verdict. Pass a spec inline to check before saving, pass intentId to check a specific saved intent, or pass nothing to resolve the repo-bound intent.md from MCP workspace roots or the launch directory. When the repository has no intent.md, or cannot be identified, it refuses rather than grading an unrelated intent; pass intentId to check a specific saved one. A failing verdict names the exact blockers; repair the fields and re-run. This is the gate that runs at preflight.pathmode.io.

Resolve an "unconfirmed" readiness dimension, in a local intent.md or in a connected workspace. Use this when check_intent_readiness reports that it READ your objective or outcomes but could not confirm them: the gate matches a fixed English vocabulary and misses valid phrasing, including other languages. confirm: state who is affected and what goes wrong for them, in your own words, and the dimension passes. waive: mark the dimension as not applicable to this change (a pure refactor has no product outcome). A waiver is a HUMAN decision, so set humanApproved only after the user has explicitly agreed. Waivers are local-mode only: a workspace will not accept one over an API key, which authenticates as a machine. The confirmation is bound to the exact text it was made against and dies automatically if that text is edited. You cannot confirm a dimension the gate read nothing for; write the field first, save, then confirm.

Get the currently active intent, preferring approved intents with real objective/outcome content over empty stubs. Returns the full IntentSpec plus an evidenceSummary (how many evidence items it cites by ID, across all trust tiers; tier counts come with include_evidence=true). Evidence is referenced by ID; pass include_evidence=true to inline the evidence content grouped by trust tier (backing, unreviewed, assumption) with per-tier counts, or call get_agent_prompt for the full evidence-backed execution prompt.

Get a single intent by ID with full details including objective, outcomes, constraints, edge cases, and relations. Evidence is referenced by ID; call get_agent_prompt for the full evidence-backed execution prompt.

Export an intent spec as .cursorrules, a CLAUDE.md or AGENTS.md section, a Claude Managed Agents Outcomes rubric, or an OpenSpec change folder. Use agents-md for Codex, Cursor, and other AGENTS.md-aware agents. Use openspec to hand the intent to an OpenSpec workflow: it writes proposal.md, a spec delta, and the verification contract in tasks.md, placed where the preflight reader finds them, and never writes into an existing change.

Adopt an existing OpenSpec change as the Pathmode intent record. Reads the change folder exactly as the preflight does (proposal.md, spec deltas, design non-goals, verification in tasks.md) and saves it as intent.md — with a workspace key, also as a workspace intent. The change folder is NEVER modified, and field text travels verbatim, so never retype fields from the change yourself. After importing, resolve any unconfirmed dimensions with confirm_intent_dimension.

Save an intent spec to intent.md in the project root — the default choice when you are working inside a repository. With a Pathmode API key set it ALSO creates or updates the intent in the workspace and records the sync in the file, so the file is the projection of the record rather than a second copy of it. Re-saving the same intent bumps its version and preserves its status; it never silently replaces a different intent. Use create_intent instead when there is no repository to write into.

List all intents in the workspace. Returns an array of IntentSpecs with their status, objectives, and metadata.

Search intents by keyword. Searches across user goals, objectives, outcomes, and constraints.