skrye-mcp (SKR-80)

A stdio MCP server that gives CC workspace-parameterized task CRUD on SKRYE, running under Row-Level Security as its own user (cc@skrye.agents). It is a thin, isolated wrapper over the existing DEFINER RPCs — no schema changes, no service_role. CC's edits go through the same RLS-checked paths a human's browser uses, so they surface in the river.

Verbs

Verb	Backing path	Surfaces in river?
list_workspaces	select workspaces (RLS)	— (discovery entry point; the only verb with no workspace param)
list_projects	select projects (RLS)	—
list_tasks	select tasks (RLS)	—
get_task	select tasks + select subtasks (RLS)	— (read includes the subtasks checklist; a read never bumps the river)
create_task	create_task_with_subtasks RPC	✅ (created activity)
set_status	set_task_status RPC	✅ (status_change)
claim_task	claim_task RPC	✅ (claimed)
add_comment	add_comment RPC	✅ (comment)
delete_task	task_soft_delete RPC	✅ (removed)
update_task	plain UPDATE (priority/description)	❌ silent — a field edit does not nudge the river (matches a human field edit, D-2)

Invariants

• Workspace is required on every verb (except list_workspaces). Resolved id-first (uuid) or by unique name. Unknown/ambiguous → hard error.
• Workspace↔task mismatch is a loud error. A task id whose project lives in another workspace is rejected (defense-in-depth atop RLS). A key ref (ABC-42) can't mismatch — it is resolved by key within the workspace.
• AAA-NNN keys are accepted anywhere a task id is, and every task payload carries its key.
• 0-row writes are hard errors. PostgREST returns 0 rows (no error) on an RLS-denied plain UPDATE; the server treats that as failure, never silent success.
• Canonical DB vocabulary (verified against live CHECK constraints): status ∈ Backlog / To Do / In Progress / Done / Canceled; priority ∈ low / normal / high. Display jargon (Waiting/Flowing/…) is a per-workspace presentation layer the MCP does not touch.

Auth

Signs in as YOUR agent (SKRYE_AGENT_EMAIL on the @skrye.agents plane) with persistSession:false + autoRefreshToken:true. JWT_EXPIRY is 3600s; the long-lived process auto-refreshes, and each call re-signs-in once if it still comes back JWT-expired. SKRYE_CC_EMAIL/SKRYE_CC_PW are accepted as legacy fallbacks (CC's original config, SKR-80).

Your agent in 5 minutes (SKR-150 — friend quickstart)

1. In Skrye, open My agents → name your agent → Create. Copy the shown-once snippet.
2. Install the server from the release tarball your operator relayed (sha256 in the release notes): npm install -g ./skrye-mcp-0.2.0.tgz
3. Paste the snippet into your Claude Code .mcp.json — it carries SKRYE_URL, SKRYE_ANON (public), and your agent's SKRYE_AGENT_EMAIL / SKRYE_AGENT_PW.
4. chmod 600 .mcp.json (or keep the password in your secret manager and reference it). Rotate the credential from My agents if you lose a device or suspect clipboard exposure — rotation kills every old session; there is no email recovery on the agent plane, by design.
5. Ask your Claude: "list my Skrye workspaces."

Config & launch (CC's own setup)

Secrets are op:// references only in .env.mcp (resolved by op run at launch; enforced by scripts/check-mcp-env-oprefs.sh in npm run preflight). Public config (SKRYE_URL, SKRYE_CC_EMAIL) lives in .mcp.json's env block. Registered in .mcp.json as:

op run --env-file=tools/skrye-mcp/.env.mcp --no-masking -- node tools/skrye-mcp/index.mjs

Targets prod (api.skrye.team) today; the server is fully env-parameterized — flip to dev by pointing the env at the dev stack, no code change.

Friend-alpha recipe (3 steps)

1. Ensure the op service account is authenticated on the host (the same one deploy/seed use).
2. In Claude Code, the skrye MCP server auto-starts from .mcp.json.
3. Ask CC: "list my Skrye workspaces" → then "in CC-dev, create a task in CLYDE titled …".

Rotation note: the CC password is read from 1Password at launch only. If you rotate Skrye-Prod-CC-Password, restart the MCP server (reload the window) for it to take effect.

Tests

npm test (node:test, offline) — resolution + every verb, including the workspace-mismatch and RLS-denied 0-row negatives. Live prod smoke test is run out-of-band via op run (not committed).