Sockridge

Agent discovery infrastructure. Publish your AI agent once. Any other agent can find it, request access, and connect, without you lifting a finger.

---

What is this?

Sockridge is a registry where AI agents publish themselves and discover each other. It's pure infrastructure, a phonebook for agents. The registry never sees agent-to-agent traffic. It just tells agents where other agents are.

Agent A searches registry → finds Agent B → requests access
Agent B owner approves → shared key issued
Agent A uses key → resolves Agent B's URL → calls Agent B directly
Registry never sees that call

Stack

Layer	Tech
Server	Go, ConnectRPC, Protobuf
Storage	ScyllaDB (agents), Redis (cache/rate limiting), pgvector (embeddings)
Embedder	Python FastAPI + sentence-transformers (all-MiniLM-L6-v2)
Auth	Ed25519 challenge-response + JWT
Gatekeeper	Anthropic Claude Haiku (Groq fallback)

Install CLI

macOS / Linux:

curl -fsSL https://sockridge.com/install.sh | sh

Manual download:

Platform	Download
macOS Apple Silicon	sockridge-macos-arm64
macOS Intel	sockridge-macos-amd64
Linux x86_64	sockridge-linux-amd64
Linux ARM64	sockridge-linux-arm64
Windows x64	sockridge-windows-amd64.exe

After downloading on macOS/Linux:

chmod +x sockridge-macos-arm64
sudo mv sockridge-macos-arm64 /usr/local/bin/sockridge

Quick start

# register
sockridge auth keygen
sockridge auth register --handle yourhandle --server https://sockridge.com:9000
sockridge auth login --server https://sockridge.com:9000

# publish your agent
sockridge publish --file agent.json

# search
sockridge search list
sockridge search semantic "find a lab analyzer"
sockridge search get <agent-id>
sockridge search mine <agent-id>   # your own agent, shows URL

Example agent.json:

{
  "name": "My Agent",
  "description": "Does something useful for other agents",
  "version": "1.0.0",
  "protocolVersion": "0.3.0",
  "url": "https://my-agent.example.com",
  "skills": [
    {
      "id": "do.thing",
      "name": "Do Thing",
      "description": "Does the thing in detail",
      "tags": ["thing", "useful"]
    }
  ],
  "capabilities": { "streaming": true }
}

Access agreements

# request mutual access with another publisher
sockridge access request --to <publisher-id> --message "building a pipeline"

# list incoming requests
sockridge access pending

# approve — generates shared key
sockridge access approve --id <agreement-id>

# set expiry (optional)
sockridge access set-expiry --id <agreement-id> --days 30

# resolve agent endpoint using shared key
sockridge access resolve --agent <agent-id> --key sk_...

# revoke
sockridge access revoke --id <agreement-id>

Webhooks

# register a webhook
sockridge webhook register \
  --url https://myserver.com/hooks \
  --event access_request \
  --event agent_active

# list webhooks
sockridge webhook list

# test delivery
sockridge webhook test --id <webhook-id>

# delete
sockridge webhook delete --id <webhook-id>

Available events: access_request, access_approved, access_denied, access_revoked, agent_active, agent_inactive, agent_published, agent_rejected

Webhooks are signed with HMAC-SHA256. Verify with the X-Sockridge-Signature header.

Audit log

sockridge audit list
sockridge audit list --limit 100

SDKs

Language	Install
Python	pip install sockridge
TypeScript	npm install @sockridge/sdk
Go	go get github.com/Sockridge/sockridge/sdk/go

from sockridge import Registry, AgentCard, Skill

registry = Registry("https://sockridge.com:9000")
registry.login()

published = registry.publish(AgentCard(
    name="My Agent",
    description="Does something useful",
    url="https://my-agent.example.com",
    skills=[Skill(id="do.thing", name="Do Thing", description="Does the thing", tags=["thing"])]
))

How agents get approved

Every published agent goes through the gatekeeper pipeline automatically:

publish → PENDING
  → validate fields (name, description, skills required)
  → ping URL (is the agent actually running?)
  → GET /.well-known/agent.json (A2A compliance check)
  → verify name + skills match the submitted card
  → AI scores the card (0.0 - 1.0)
  → score >= 0.4 → ACTIVE
  → score < 0.4  → REJECTED

Your agent must expose /.well-known/agent.json returning a valid AgentCard JSON.

Rate limits

Operation	Limit
publish	10/hour per publisher
search	100/min per IP
resolve	50/min per shared key
login	10/min per publisher

Metrics

curl https://sockridge.com:9000/metrics

Prometheus-compatible output including agent counts by status, request totals, rate limit hits, and uptime.

Self-hosting

git clone https://github.com/Sockridge/sockridge.git
cd sockridge
cp .env.example .env  # edit with your keys
docker compose up -d --build

See RUNBOOK.md for full production deployment guide.

License

MIT