generated: '2026-08-17'
method: generated
source:
- https://learn.elium.com/reference/
- https://learn.elium.com/en/api/
- https://help.elium.com/en/articles/12767574-mcp
note: >-
  Elium publishes no OpenAPI, so these skills are grounded in its published GraphQL reference instead. Every
  operation named in every skill was checked against learn.elium.com/reference/ and exists there verbatim;
  the five operations Elium marks Deprecated (statusStory, storyApprovalEdit, storyMigrate, userFollow,
  userPresenceUpdated) are called out as such and never recommended. No provider-published Agent Skills or
  AGENTS.md file exists to save verbatim - searched the docs, help centre and llms.txt.
grounded_in: graphql
skill_count: 3
skills:
- file: elium-search-knowledge-base.md
  name: Search the Elium knowledge base and read an article
  api: graphql/elium-graphql-operations.yml
  surfaces: [graphql, mcp]
  operations: [me.answer.semanticSearch, me.search, story, searchFeedback]
  access: read
- file: elium-publish-a-story.md
  name: Publish and update an Elium story
  api: graphql/elium-graphql-operations.yml
  surfaces: [graphql]
  operations: [storyPost, storyEdit, storySpace, storyApprovalRequest, storyApprovalStatus, storyExpire,
    storyLock, storyUnlock, story]
  access: write
- file: elium-connect-an-ai-assistant.md
  name: Connect an AI assistant to Elium over MCP
  api: mcp/elium-mcp.yml
  surfaces: [mcp, oauth]
  operations: [FindRelevantArticles, FindSpecificArticles, ReadFullArticle, connectedAppCreate]
  access: read
cross_cutting_rules_every_skill_carries:
- The endpoint is per tenant - https://{platform}.elium.com/... There is no shared vendor host and a tenant
  name must never be guessed.
- The API and the MCP server are Enterprise-plan only; Team and Corporate cannot call them.
- OAuth 2.0 bearer, one scope (apiv1), 3600-second access tokens, no API keys.
- HTTP 200 is not success - handled errors are in-band (top-level GraphQL errors, or
  data.<mutation>.errors[] with __typename and path).
- There is NO idempotency mechanism, so no write may be blind-retried; verify the effect first.
- No rate limits and no rate-limit response headers are published - throttling is unobservable.
- Everything is evaluated with the token user's own permissions, so an empty result may mean "not
  permitted" rather than "not found".