Features
The Tool Registry That Works Even When Tools Are Missing
Not just a directory — a self-healing discovery layer with hybrid retrieval, 3-tier fallback, and multi-protocol output.
Self-Healing Registry
Never Returns Empty — Auto-Crawls GitHub When Tools Are Missing
When no indexed tool scores above the quality threshold, ServAgent automatically searches GitHub for new repositories matching your task, indexes the best result, and returns it — all in the same request. The registry grows with every query.
- Tier 1: Hybrid BM25 + neural embedding search across 6,000+ indexed tools
- Tier 2: Live GitHub gap-fill crawler (triggers when no tool scores above threshold)
- Tier 3: Generates a working tool from free public APIs as a last resort
- Results are permanently cached — repeated queries get instant responses
Recommendation Accuracy
Hybrid Retrieval — BM25 + Neural + LLM Reranking
ServAgent uses a three-stage retrieval pipeline: BM25 lexical matching, multilingual neural embeddings (e5-large, 1024-dim), and LLM-based reranking (DeepSeek). Reciprocal Rank Fusion (RRF) combines lexical and semantic signals. Cosine calibration ensures scores reflect true relevance. All metrics come from reproducible eval scripts included in the repo.
- BM25+ multi-field scoring (name x4, category x3, tags x2.5, description x1) with query expansion
- Neural embeddings: multilingual-e5-large (1024-dim) for cross-lingual semantic matching
- LLM reranker: DeepSeek re-orders top-10 candidates by task relevance (~$0.001/query)
- Eval: 50% top-1, 63% top-3 accuracy on 1,119-case benchmark (NDCG@5 = 0.60)
Multi-Protocol
One Registry, Every Agent Framework
ServAgent converts tool definitions on the fly. Fetch a tool in OpenAI function-calling format, Anthropic tool-use schema, Google Agent protocol, or MCP JSON-RPC — whichever your agent needs, in one API call.
- POST /deploy/batch — convert up to 50 tools in one request
- GET /deploy/:id/:protocol — single tool, any protocol
- POST /api/mcp — full MCP JSON-RPC 2.0 gateway (tools/list, tools/call)
- GET /api/v1/openapi.json — OpenAPI 3.1 spec for ChatGPT GPT Actions
Smart Pipeline
Pipeline Orchestrator — Caching, Compression, Feedback Loop
The pipeline layer sits between your agent and the tool registry, adding multi-level caching, response compression, and a feedback signal so the system learns which tools actually work for each task.
- POST /pipeline/resolve — single endpoint for query -> ranked tools -> protocol-converted schemas
- POST /pipeline/feedback — record success/failure to improve future rankings
- GET /pipeline/stats — live cache hit rate, p50/p95 latency, token savings
- Shadow SDK (POST /telemetry/ingest) — zero-overhead event collection from your agent
ServAgent vs Smithery
Based on publicly available information, April 2026
| Feature | ServAgent | Smithery |
|---|---|---|
| Tool count | 6,000+ (auto-growing) | 3,000+ (community submit) |
| Recommendation accuracy | 63% top-3 (1,119-case eval) | Not published |
| Cold-start / missing tools | GitHub gap-fill + generate | Wait for community PR |
| Multi-protocol output | MCP / OpenAI / Anthropic / Google | MCP primarily |
| Provider analytics | API + Dashboard (Pro) | None |
| Self-hosted option | Open source | Closed |