MCP stdio server for Axon — exposes the Axon REST API as MCP tools so Copilot (or any other agent) can call them from agent mode.
Axon RAG MCP server has well-structured tool definitions with consistent naming (verb_noun pattern), comprehensive descriptions (100-300 chars typical), and explicit input schemas. However, output schemas are not documented, error handling lacks recovery guidance, and several tools expose internal state mutations (switch_project) without safeguards. Tool descriptions are LLM-optimized but some parameters lack format constraints. No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels. Composition is sound, tools chain well via project/doc_id references.
Create a new knowledge base project.
Remove documents from the knowledge base by their IDs. Deletes from both the vector store and the BM25 index.
Poll the status of an async ingest job started by ingest_path. Returns a dict with: job_id, status (processing|completed|failed), started_at, completed_at, path, error.
Return documents that have not been re-ingested within *days* calendar days. Use this to identify outdated knowledge that should be refreshed. Only documents ingested during the current server process lifetime are tracked — restart tracking begins fresh after each server restart.
Ingest a local file or directory into the knowledge base (async). Returns immediately with a job_id. Poll get_job_status(job_id) until status is 'completed' or 'failed'. Path must be within RAG_INGEST_BASE.
Output schemas not documented. Tools return structured data (job_id, status, chunks, etc.) but LLMs cannot plan downstream calls without knowing response field names and types. Violates pattern:tool requirement.
switch_project mutates global server state and warns 'Do not call from concurrent request handlers.' This violates stateless request handling (current MCP spec). No confirmation step for destructive operation. Agents cannot safely use this tool.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels. delete_documents is DESTRUCTIVE, ingest_* are WRITE, search/query are READ_ONLY. LLMs cannot reason about safety without explicit hints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 66 | 2026-07-28+ | v2 |
Ingest a single text document into the Axon knowledge base. Prefer ingest_texts for multiple documents — it uses one embedding call. Always set metadata.source so the collection can be audited. Duplicate content (same SHA-256) is silently skipped; status will be 'skipped'.
Ingest multiple documents in a single batched embedding call. Each item must have at least a "text" key. Optional keys: "doc_id", "metadata". This is the preferred ingest tool — never call ingest_text in a loop.
Fetch an HTTP/HTTPS URL and ingest its text content. HTML is stripped automatically. Private/internal URLs (127.x, 10.x, 192.168.x, 169.254.x, 172.16-31.x) are blocked server-side.
List all indexed sources in the active project with chunk counts. Call this before a large ingest to check what's already indexed and avoid re-ingesting duplicate content.
List all knowledge base projects. Returns on-disk projects (with metadata) plus any project seen only in the current server session. Call this to discover available namespaces before switching or querying a project.
Ask a question and get a synthesised answer from the knowledge base. Performs retrieval + generation in one call. Use search_knowledge instead if you need to inspect raw chunks before answering.
Re-ingest all sources that have changed on disk since they were last indexed. Compares SHA-256 hashes of previously indexed files against their current on-disk content. Changed files are re-chunked and re-embedded; unchanged files are skipped. Returns a job_id for async polling via get_job_status.
Retrieve raw document chunks from the knowledge base. Best for multi-step reasoning where you want to inspect individual chunks before synthesising an answer. Use query_knowledge for direct answers.
Switch the knowledge base to a different project. WARNING: This mutates global server state. Do not call from concurrent request handlers. Prefer passing 'project' directly to ingest tools instead.
Error handling lacks recovery guidance. No tool description explains what to do if ingest_path fails, if search_knowledge returns no results, or if project switch fails. Agents cannot self-correct.
Parameter constraints not fully specified. 'top_k' lacks min/max bounds (should be 1-100+). 'days' in get_stale_docs lacks range. 'filters' object structure undefined. LLMs may pass invalid values.