API Server & MCP Protocol

Axum server for document extraction, plus the rmcp Model Context Protocol surface

Locations: crates/xberg/src/api/ (router.rs, handlers.rs, startup.rs, types.rs, error.rs, jobs.rs) and crates/xberg/src/mcp/. There is no api/server.rs.

REST routes

Registered in api/router.rs, all on one Router:

There is no POST /extract-url (URL ingestion is a field on ExtractInput passed to /extract) and no POST /batch (batch is /extract-async + /jobs/{job_id}).

Middleware, in order: DefaultBodyLimit::max(limits.max_request_body_bytes) + RequestBodyLimitLayer (default 100 MB), CORS, request-id, compression, catch-panic, sensitive-header stripping, tracing. CORS is built explicitly as CorsLayer::new().allow_origin(Any).allow_methods(Any).allow_headers(Any) and warns loudly; restrict it with XBERG_CORS_ORIGINS.

Caching

crates/xberg/src/cache/ — GenericCache is a filesystem-backed store with LRU-style eviction, not an in-memory map. Keys are BLAKE3 content hashes (blake3_hash_bytes / blake3_hash_file, cache/utilities.rs). Eviction is bounded by max_age_days, max_cache_size_mb and min_free_space_mb — there is no entry-count limit.

Error handling

ApiError is a struct, not an enum: { status: StatusCode, body: ErrorResponse } (api/error.rs). Status comes from the constructor, not a variant:

• validation() → 400, unprocessable() → 422, internal() → 500, bad_gateway() → 502
• From<XbergError> picks one via error.api_status_category() (Validation / Unprocessable / Internal)

There is no 404, 413 or 503 path with a named variant. Do not match on ApiError.

MCP

crates/xberg/src/mcp/. Transport is a single nested rmcp streamable-HTTP service — Router::new().nest_service("/mcp", http_service) — not a set of /mcp/* REST paths. Tools, resources and prompts are JSON-RPC methods on it. Stdio transport serves the same router over stdin/stdout.

Tools (9)

extract, extract_batch, detect_mime_type, list_formats, cache_stats, cache_clear, get_version, cache_manifest, cache_warm. The set is pinned by test_all_tools_are_registered in mcp/server.rs. There is no get_capabilities.

extract, extract_batch and cache_warm are task-eligible (TASK_ELIGIBLE_TOOLS, mcp/server.rs).

Resources

mcp/resources.rs: xberg://formats, xberg://models, xberg://languages/ocr, plus xberg://presets/embeddings behind #[cfg(feature = "embeddings")].

Prompts (3)

mcp/prompts.rs: extract_document, extract_with_ocr, semantic_search.

Environment variables

There is no .env.example. Server-side vars are read in core/server_config/env.rs:

• XBERG_HOST, XBERG_PORT (defaults 127.0.0.1:8000)
• XBERG_MAX_REQUEST_BODY_BYTES, XBERG_MAX_MULTIPART_FIELD_BYTES (both default 100 MB)
• XBERG_CORS_ORIGINS (comma-separated)

Extraction-side vars are documented on ExtractionConfig::apply_env_overrides (core/config/extraction/env.rs) — XBERG_OCR_BACKEND, XBERG_OCR_LANGUAGE, XBERG_CHUNKING_MAX_CHARS, XBERG_CACHE_ENABLED, XBERG_LLM_*, and others. Read that doc comment rather than guessing a name.

Critical Rules

REST

1. Validate uploads — MIME type, size, magic bytes; never trust the filename.
2. Size limits are configurable — always read limits.max_request_body_bytes, never hardcode.
3. Errors must be actionable — include the operation and a remediation hint in ErrorResponse.
4. CORS is permissive by default — production deployments must set XBERG_CORS_ORIGINS.
5. Long work goes to /extract-async — do not block a request thread on a multi-minute extraction.

MCP

1. Register a new tool in mcp/server.rs and extend test_all_tools_are_registered — the test is the contract.
2. Feature-gate resources the same way xberg://presets/embeddings is — a missing feature must not break resources/list.
3. Resources are static — no network or filesystem scans in a resource handler.
4. Tools need timeouts — a hung tool blocks the agent.

Related Skills

• extraction-pipeline-patterns — the core extraction the handlers and tools call
• chunking-embeddings — optional chunking/embedding parameters
• config-loading-precedence — server-mode precedence and env overrides