Repository avatar
AI Tools
v1.0.1
active

evermuse

com.evermuse/evermuse

Search your customer interviews, calls, feedback, competitor intel and roadmap from any AI agent.

Documentation

Evermuse for Claude Code and Cursor

Make your AI agent speak with the voice of your customers.

Evermuse is a plugin for Claude Code and Cursor that turns everyday product work — specs, dev plans, research, roadmaps, gap analyses, PR reviews — into deliverables grounded in real customer evidence. Instead of plausible-sounding generic answers, you get verbatim quotes, real needs, and actual pain points, each with a source link back to the conversation it came from. And the important things you produce get saved back into Evermuse, so your knowledge base compounds.

It combines proven PM/research methods (adapted from pm-skills) and spec-driven-development discipline (adapted from spec-kit) with the Evermuse MCP, which searches your recorded calls, interviews, and feedback.

What changes after you install it

Ask "what do customers think about our onboarding?" and instead of a guess you get:

Short answer: Onboarding friction centers on manual data entry — 7 mentions across 5 accounts, mostly frustrated.

Theme 1 — Re-keying data by hand (7 mentions / 5 accounts)

"We lose half a day every week re-keying this into the spreadsheet." — Dana K., Acme onboarding call, 12 May 2025 · View in Evermuse [^1]

Every spec gets a Customer Evidence section. Every dev plan opens with the customer quotes that justify it. Every PR review checks the diff against what customers literally asked for. That's the difference.

Install

Claude Code

# Add this repo as a plugin marketplace, then install the plugin
claude plugin marketplace add Usermuse/em-plugin
claude plugin install evermuse@evermuse

Or point at a local checkout during development:

claude plugin marketplace add /path/to/em-plugin
claude plugin install evermuse@evermuse

Cursor

Install from the Cursor Marketplace, or add just the MCP server with a one-click deeplink:

Add Evermuse to Cursor

The Cursor plugin ships the skills and the MCP server.

How invocation differs between the two clients. In Claude Code, skills auto-activate — you describe the task in plain language and the right skill fires. Cursor works differently: skills are invoked explicitly, by typing /skill-name in chat or @skill-name to attach one as context. So in Cursor the skill name is the command:

Claude CodeCursor
/evermuse:spec <feature>/write-feature-spec
/evermuse:research <question>/customer-research
"write a spec for bulk CSV export" (auto-activates)/write-feature-spec (explicit)

Because Cursor won't auto-fire a skill, the short aliases matter more there, not less — so Cursor gets its own command layer in cursor-commands/:

/spec        /dev-plan     /research     /setup      /save
/discover    /brainstorm   /pricing      /strategy   /plan-launch    …28 in total

Each carries Cursor's name / description frontmatter. The Claude versions in commands/ use a different frontmatter shape plus $ARGUMENTS, which Cursor does not support, so the two directories are kept separate and each client reads its own.

Six Claude commands are deliberately not duplicated — business-model, gap-analysis, release-notes, review-pr, test-scenarios and value-proposition share a name with a skill that already does the same job, so in Cursor you invoke those as /gap-analysis etc. directly and a duplicate command would only make the / list ambiguous.

Authenticate the MCP

The plugin bundles the Evermuse MCP server (https://api.evermuse.com/api/mcp). Before the skills can ground anything, connect it:

  • Claude Code — OAuth (recommended): run /mcp in an interactive session and authorize evermuse.
  • Cursor — OAuth: open Settings → MCP and authorize evermuse. Cursor registers itself dynamically (DCR), so there's no client ID to configure.
  • API key (either client): add an Authorization header for the evermuse server in your MCP settings.

Then run /evermuse:setup to pick your Product and confirm grounding works. (A Product is required for every Evermuse call; a Project like Discovery/Support/Sales is optional and only needed when you're focused on one research effort.)

Quickstart

/evermuse:setup                      # connect + pick your product
/evermuse:research what do customers think about pricing?
/evermuse:spec bulk CSV export for the reporting page
/evermuse:dev-plan specs/bulk-csv-export/spec.md
/evermuse:review-pr 482

The doctrine (how grounding works)

Every skill follows Ground → Work → Cite → Save:

  1. Ground — verify the right Product, then run 2–4 differently-worded searches (declaring a nature: evidence = customer voice, context = market, guidance = company strategy) plus verbatim quotes.
  2. Work — apply the skill's method, shaped by what the evidence says.
  3. Cite — every customer-derived claim carries a source badge and a [^n] footnote.
  4. Save — offer to persist the deliverable back into Evermuse via add_source.

The full doctrine lives in the using-evermuse skill, which every other skill defers to.

The MCP tools

The submitted default cloud OAuth configuration exposes these 27 tools:

GroupTools
Read-only: guidancefind_skills · read_skills
Read-only: product contextget_products · get_product_summary · get_projects
Read-only: workflow (start here)customer_research · create_prd · write_brief · write_feature_spec · user_personas · user_stories · competitor_analysis · summarize_conversation
Read-only: evidence and sourcessearch · view_item · find_sources · read_source
Read-only: internal assetslist_competitors · get_competitor_capabilities · get_opportunities · get_shaping_notes · read_shaping_note
Write: additiveadd_source · add_signals · create_shaping_note
Write: overwrites existing data (destructive)update_signals · update_shaping_note

Workflow tools are read-only despite their imperative names: create_prd creates nothing — you write the document, and it is stored only if you then call a write tool.

Not every workspace gets exactly this list. Workspaces with shaping writes disabled don't get create_shaping_note / update_shaping_note, local-storage workspaces get only the guidance tools from the cloud endpoint, and other feature flags can add tools. tools/list is the authority for what a given workspace is served.

A workflow tool runs the job's first grounding step and returns the methodology, so start from one when it matches the task rather than reaching straight for search. Every product-scoped call takes product_id as a parameter — there is no session to switch.

Tool annotations

Every tool sets all four MCP tool annotations explicitly to true or false, plus a title.

Tool(s)readOnlyHintdestructiveHintidempotentHintopenWorldHintWhy
The 22 read-only tools: find_skills · read_skills · get_products · get_product_summary · get_projects · customer_research · create_prd · write_brief · write_feature_spec · user_personas · user_stories · competitor_analysis · summarize_conversation · search · view_item · find_sources · read_source · list_competitors · get_competitor_capabilities · get_opportunities · get_shaping_notes · read_shaping_notetruefalsetruefalseRetrieve guidance or authorized workspace data. They never create, overwrite, delete or publish the user's sources, signals or notes.
add_sourcefalsefalsefalsetrueAdds a source to the caller's workspace and queues ingestion; the response confirms submission, not completion. Meeting imports can fetch a caller-supplied public HTTPS media URL (media.url), which is why it is open-world. Meeting and communication imports deduplicate on external_source + external_id, but note, document and spreadsheet modes don't, so a repeated call can create another source. It never overwrites an existing source.
add_signalsfalsefalsetruefalseAdds up to 100 signals to an existing source. Replays are deduplicated on source, type and content, so repeating a call creates nothing new.
create_shaping_notefalsefalsefalsefalseCreates a new shaping note and never modifies an existing one. Repeating the call creates a second note.
update_signalsfalsetruefalsefalseOverwrites the passed fields on up to 100 existing signals (null clears an optional title or attribution). Signals have no version history. Repeating a title, content or type update advances the embedding revision and queues embedding work again, so a repeat has further effects.
update_shaping_notefalsetruefalsefalseReplaces each passed field outright, content and tags included. Every successful update, a repeat included, creates a new version, so it is not idempotent. Version history allows recovery, but the current fields are still overwritten.

Idempotency here means a repeated identical call has no further effect, not that the data can't change between calls. Every tool except add_source works only on Evermuse guidance or the caller's own workspace data (hosted storage and embedding services included); add_source can also download a media file from a URL the caller supplies. The third-party proxy tools (find_tool / call_tool) are not served to OAuth clients. No tool on this surface renders UI.

Commands

Cursor — 28 plain-Markdown commands in cursor-commands/. Same short names as below, without the evermuse: prefix (/spec, /dev-plan, /research). The six that duplicate a skill name are omitted; invoke those as /gap-analysis, /review-pr etc. directly.

Claude Code — the /evermuse:* commands in commands/, listed here:

Flagship: /evermuse:setup · spec · dev-plan · gap-analysis · review-pr · research · release-notes · save

Discovery & execution: discover · brainstorm · triage-requests · interview · write-prd · write-stories · plan-okrs · transform-roadmap · test-scenarios · pre-mortem · red-team-prd · meeting-notes · ship-check

Strategy, research & GTM: strategy · business-model · value-proposition · market-scan · pricing · research-users · competitive-analysis · analyze-feedback · plan-launch · growth-strategy · battlecard · market-product · north-star

Skills also auto-activate when you describe the task in plain language ("write a spec for…", "what are customers saying about…", "review this PR") — you don't have to remember command names.

Skills catalog

  • Foundation: using-evermuse (the doctrine + MCP mechanics every skill uses)
  • Flagships: write-feature-spec, development-plan, gap-analysis, review-pr, customer-research
  • Discovery: brainstorm-ideas, deep-dive, assumptions, prioritize-features, analyze-feature-requests, opportunity-solution-tree, interview-script, summarize-conversation
  • Strategy: product-strategy, product-vision, value-proposition, business-model, pricing-strategy, strategy-frameworks, strategy-red-team
  • Execution: create-prd, user-stories, brainstorm-okrs, outcome-roadmap, release-notes, test-scenarios, shipping-artifacts, daily-brief
  • Market research: user-personas, segmentation, customer-journey-map, market-sizing, competitor-analysis, sentiment-analysis
  • GTM & growth: gtm-strategy, beachhead-segment, ideal-customer-profile, competitive-battlecard, positioning-and-messaging, north-star-metric
  • Onboarding & meta: setup-evermuse, initial-report, list-capabilities, schedule-new-ingestion-tasks

skills/ is generated from the Evermuse monorepo (yarn sync:em-plugin). Do not edit it here — open an issue instead, and the change is made upstream and synced down.

Tips

  • The plugin works best when your project's CLAUDE.md mentions the product name — it helps the agent pick the right Evermuse Product automatically. (Optional.)
  • If Evermuse isn't connected, skills still run but produce ⚠ ungrounded output and tell you to authorize the MCP — they never fake evidence.
  • Secondary Evermuse assets (the AI-generated roadmap, shaping notes, competitor list, research questions) are used sparingly and always labeled — customer conversations are the source of truth.

Contributing

  • skills/ is generated. Request changes upstream in the Evermuse monorepo; a sync PR brings them here. Commands, manifests, the README and the logo are hand-written and edited here.
  • The setup header inlines logomark.svg. The setup-evermuse skill embeds the SVG markup directly rather than loading it from a URL, so if the logomark changes, update the inline copy in that skill too.
  • Bump the version before merging a sync PR. Cursor re-reviews every update, so each release needs version raised in .claude-plugin/plugin.json, .cursor-plugin/plugin.json, .codex-plugin/plugin.json and .claude-plugin/marketplace.json.

Credits

Built on phuryn/pm-skills (MIT) and github/spec-kit (MIT). See ATTRIBUTION.md. Licensed MIT — see LICENSE.

No installation packages available.