Editor agents (MCP): ask your editor what broke
Route: /mcp (sidebar: Editor agents, under Connect)
Wire Mushi Mushi into Cursor, Claude Code, Windsurf, or any MCP client. Once connected, your agent can read reports, pull fix context, and optionally dispatch fixes — without leaving the editor.
The page has four tabs: Summary, Setup, Catalog and Examples. The Catalog tab mirrors the server registry (130 tools, 8 resources, 4 prompts).
Quickstart
The Setup tab opens with a 3-step strip, Get an agent talking in 60 seconds. Step 1
shows complete when you already have an mcp:read key.
1. Generate an API key
Mint a key with MCP read (to browse) or MCP read + write (to dispatch fixes), or use Add to Cursor, which creates a key for you. This is separate from your app’s SDK ingest key (report:write).
2. Paste the snippet
Copy the Stdio (.cursor/mcp.json) or Hosted HTTP block, or click Add to Cursor (or the button for your editor on Connect) for a one-click deeplink. The deeplink includes MUSHI_FEATURES=triage,fixes,inventory,setup,docs (lean catalog).
3. Ask the agent
Restart your IDE, then ask: “list mushi tools” or “call diagnose_setup”. See MCP quickstart and Incident loop.
Install snippets
.cursor/mcp.json (stdio — recommended)
{
"mcpServers": {
"mushi-myproject": {
"command": "npx",
"args": ["-y", "@mushi-mushi/mcp@0.30.4"],
"env": {
"MUSHI_API_ENDPOINT": "https://<your-ref>.supabase.co/functions/v1/api",
"MUSHI_PROJECT_ID": "<your-project-id>",
"MUSHI_API_KEY": "mushi_...",
"MUSHI_FEATURES": "triage,fixes,inventory,setup,docs"
}
}
}
}There is no --project-id CLI flag on @mushi-mushi/mcp. Pass the project via MUSHI_PROJECT_ID in env. Server names use mushi-<slug> (admin deeplink) or bare mushi for legacy single-project configs.
Hosted HTTP (no local subprocess)
{
"mcpServers": {
"mushi-http": {
"type": "http",
"url": "https://<your-ref>.supabase.co/functions/v1/mcp?features=triage,fixes,inventory,setup,docs",
"headers": {
"X-Mushi-Api-Key": "mushi_...",
"X-Mushi-Project-Id": "<your-project-id>"
}
}
}
}Full catalog: append ?features=all (or set MUSHI_FEATURES=all for stdio).
Smallest set: ?features=vibe lists a dozen tools for the reported-bug loop (reports,
diagnosis, fix context, mark fixed, reply, setup, docs); chat connectors such as
Lovable use it (guide).
Key status
Two badges next to the strip show whether you have:
mcp:read— required for read tools (reports, fix context, lessons, docs search)mcp:write— required for dispatch, merge, QA triggers, and other mutations
Write tools mutate production data. Store mcp:write keys in IDE config or CI secrets — never in a frontend bundle that ships to end users.
Tool catalogue (representative)
See the admin Catalog tab for the full list. Names below match tools/call exactly.
Read tools (mcp:read)
| Tool | Description |
|---|---|
get_recent_reports | List recent reports with status/category/severity filters |
get_report_detail | Full report payload — logs, screenshot, classification |
get_fix_context | One-shot fix brief — root cause, repro, blast radius |
get_similar_bugs | pgvector nearest-neighbour duplicates |
search_reports | Semantic + keyword search |
diagnose_setup | Unified ingest + MCP + dispatch diagnosis |
search_mushi_docs | Keyword search over Mushi Mushi documentation |
query_lessons / list_lessons | Evolution-loop learning rules |
get_inventory / diff_inventory | Inventory v2 snapshot and gate diff |
Write tools (mcp:write)
| Tool | Description |
|---|---|
dispatch_fix | Dispatch a triaged report to the fix-worker (opens draft PR when configured) |
merge_fix | Squash-merge the fix PR and mark report Fixed |
transition_status | Move a report through triage states |
run_qa_story | Trigger a QA Coverage story run |
start_skill_pipeline | Attach a skill pipeline to a report |
Resources
| URI | Description |
|---|---|
project://dashboard | PDCA dashboard snapshot |
project://stats | Report counts and severity breakdown |
mushi://activation | Setup posture — SDK heartbeat, MCP, next action |
inventory://current | Current inventory snapshot with gate status |
evolution://history | Judge scores and lesson inductions (30 days) |
privacy://status | BYOK / retention / region posture |
Prompts
MCP prompts orchestrate multiple tools into a single agent turn:
| Prompt | Description |
|---|---|
summarize_report_for_fix | Build a fix plan from get_fix_context + blast radius before coding |
triage_next_steps | Five-item priority list from dashboard + queue |
explain_judge_result | Ship / iterate / dismiss from judge scores |
mushi_setup | Diagnose blocked setup via activation + integration health |
Use cases
- Incident loop —
get_fix_context→summarize_report_for_fix→ paste into Cursor (guide) - Morning triage —
get_recent_reports+triage_next_steps - PR-time lessons —
query_lessonsbefore touching files - Setup unblock —
diagnose_setupwhen reports don’t appear - Automated dispatch — CI calls
dispatch_fixon critical reports (requiresmcp:write)
Related pages
- Manage apps: mint MCP API keys
- Connect: unified SDK + MCP install hub
- MCP OAuth consent: sign in from the editor instead of pasting a key
- SDKs → MCP:
@mushi-mushi/mcppackage reference - MCP tools reference: full catalog from
catalog.ts - Fix orchestrator: what happens after
dispatch_fix