Agent Tool Registry

Validated agent tools with reliability filtering.

MCP servers, OpenAPI tools, and AI plugins — continuously tested, scored on rolling reliability metrics, and surfaced only when they pass. Filter by uptime, latency, auth type, or capability and get exactly what your agent can depend on.

Query the registry How it works
—
Registry Entries
—
Open
—
Auth Gated
—
Production Safe
How it works

Four pillars.
One registry.

01
Discovery
Continuously indexes GitHub, npm, PyPI, Smithery, well-known MCP endpoints, OpenAPI specs, and /.well-known/ai-plugin.json manifests. New tools enter a candidate queue the moment they appear.
02
Validation
MCP candidates are tested via live handshake — protocol negotiation, tools/list enumeration, latency measurement — under a 10-second hard timeout. OpenAPI and AI plugin entries are catalogued directly. State transitions automatically based on results.
03
Trust metrics
Rolling uptime (24h, 7d), handshake success rate, and latency percentiles (p50/p95) — computed continuously from validation history. No trust score without evidence.
04
Filtering API
Query by production_safe, min_uptime, max_latency_p95, auth_type, or read_only. Active server responses include a ready-to-paste MCP config snippet.
Live API

Query the registry
right now.

The Arclan registry exposes a simple REST API. Filter by reliability thresholds, auth type, or capability — and run any query live below.

GET /api/agents/default-mcps Ready-to-use defaults
GET /api/registry List servers
GET /api/registry?production_safe=true Production-safe with metrics
GET /api/stats Global counts
POST /api/submit Submit a server
GET /api/agents/default-mcps 
# Default production-safe MCPs — trust, flags, and decision included
$ curl -s https://arclan.ai/api/agents/default-mcps \
    -G \
    --data-urlencode "limit=10"
200 OK 

          

        

      

    

  







  

    
Trust model

    
Not a directory.
A validated registry.

    

      

        

          Existing directories are static lists — human-curated, not machine-evaluated. Arclan continuously validates reachable MCP servers, catalogues OpenAPI tools and AI plugins, and scores everything from observed behaviour. Nothing is marked healthy unless the checks support it.
        

        

          

            
→

            

              
Composite score

              
70% weight on the most recent validation result, 30% on the 30-day rolling history. Score range 0–100.

            

          

          

            
→

            

              
State machine

              
Every entry lands in one of eleven states: active (connectable MCP), openapi (REST spec), ai_plugin (ChatGPT plugin manifest), auth gated, degraded, partial, stdio, unreachable, invalid, or zombie. No manual review involved.

            

          

          

            
→

            

              
Reliability metrics

              
Rolling uptime (24h, 7d), handshake success rate, and latency percentiles (p50/p95) — computed continuously from validation history. Filter with production_safe=true, min_uptime, or max_latency_p95.

            

          

        

      


      
      

        
Live reliability samples

        

          
Loading live scores…