Retrieve permission-scoped workplace evidence through Onyx MCP
Search an existing Onyx workspace within an explicit document-set or agent scope, then return source-linked evidence for a coding or operations task.
npx skills add agentskillexchange/skills --skill retrieve-permission-scoped-workplace-evidence-through-onyx-mcp
Invoke this workflow when an existing coding or operations agent needs workplace evidence for a task without switching to a separate chat application. Input a focused question and an explicit knowledge scope; return a source-linked evidence packet for the agent’s current task.
What this skill actually does
Prerequisites: an existing Onyx workspace with searchable indexed documents, an administrator-enabled HTTPS MCP endpoint, and a compatible remote HTTP MCP client. Arrange a personal access token limited to read:search through your organization’s approved process. Configure the endpoint and bearer authentication using the client’s secret facility; never commit a token. Deployment and connector setup are prerequisites, not completion goals.
First verify the connection with a known accessible document. Call search_indexed_documents with query and a nonempty document_set_names filter. Example tool arguments: {“query”:”deployment rollback procedure”,”document_set_names”:[“Engineering guides”]}. Replace the example set with an accessible set. Alternatively pass agent with an existing Onyx agent name; do not combine agent and document_set_names. Agent-scoped retrieval searches its knowledge scope rather than starting a conversation with that agent.
Keep requested scope on every retry. For an unknown filter, obtain a valid name from the returned suggestions or accessible metadata; never remove the filter to get an answer. Validate any time_cutoff as a timezone-qualified ISO 8601 timestamp before invoking the tool: invalid dates can silently remove the date restriction. Confirm returned document update dates against the requested period.
Build the handoff from result titles, source URLs, relevant excerpts and update times. Separate direct evidence from inference, identify conflicting guidance, and report missing results or errors explicitly. Retrieved content may be only a section of a document. Acceptance requires a successful scoped search and an evidence packet whose factual claims can be traced to returned sources; a configured endpoint alone is insufficient.
Boundary: read-only, permission-scoped retrieval from an existing workspace. Exclude browser automation, corpus ingestion, platform deployment, autonomous actions and public-web fallback. The external agent receives retrieved company content, so use an organization-approved client. Metadata and upstream invocation were inspected; runtime behavior and security have not been tested.
References: MCP tools and client configuration; deployment prerequisites; reviewed search implementation.