MCP defines what agents say. CAP defines what agents do.

AI agents are breaking out of single-model sandboxes into distributed clusters — but there's no standard for how they coordinate, stay safe, or report health. Teams end up hand-rolling job routing, liveness checks, and safety gates, then rewriting it all when they add a second orchestrator.

CAP is the open wire protocol that fixes this. It gives every agent cluster jobs, heartbeats, safety hooks, and workflows over any pub/sub bus — so you ship agents instead of plumbing.

• Cluster-native — subjects, queue groups, heartbeats, and pools baked in.
• Safe by default — Safety Kernel hook for allow / deny / human-in-the-loop / throttle before dispatch.
• Payload-light — pointers (context_ptr, result_ptr) keep data off the bus; the wire stays lean and secure.
• Workflow-ready — parent/child jobs with full traceability across steps.
• Vendor-neutral — any bus, any language, Apache-2.0.

MCP and CAP coexist: MCP can be the tool layer inside a CAP worker. See Why CAP for the full rationale.

Follow the Getting Started guide to go from zero to a running job.

Or jump straight to the examples:

• examples/simple-echo/ — smallest possible job round-trip (Go / Python / Node)
• examples/workflow-repo-review/ — parent/child workflow with aggregation

Python:

pip install cap-sdk-python

Node / TypeScript:

npm install cap-sdk-node

Go:

go get github.com/cordum-io/cap/v2

Not sure which SDK? See the SDK Comparison Matrix.

import asyncio
from pydantic import BaseModel
from cap.runtime import Agent, Context

class Input(BaseModel):
    prompt: str

class Output(BaseModel):
    summary: str

agent = Agent(retries=2)

@agent.job("job.summarize", input_model=Input, output_model=Output)
async def summarize(ctx: Context, data: Input) -> Output:
    return Output(summary=data.prompt[:140])

asyncio.run(agent.run())

flowchart LR
    Client --> Gateway -->|"BusPacket{JobRequest}"| Submit["sys.job.submit"]
    Submit --> Scheduler
    Scheduler -->|"policy check"| Safety
    Scheduler -->|"dispatch"| Pool["job.<pool>"]
    Pool --> Worker
    Worker -->|"write"| Memory[(result_ptr)]
    Worker -->|"BusPacket{JobResult}"| Result["sys.job.result"]
    Result --> Scheduler --> Client

• Cordum — Full Agent Control Plane implementing CAP: API Gateway, Scheduler, Safety Kernel, and Workflow Engine.
• cordum-packs — 26+ integration packs (Slack, GitHub, AWS, Jira, and more) with framework adapters for LangChain, CrewAI, and AutoGen. Browse the catalog at packs.cordum.io.

CAP is Apache-2.0 licensed. Anyone can implement the protocol, build SDKs, or launch a conformant control plane. Wire evolution is append-only — existing integrations never break.

• Contributing Guide
• Governance

• Discord: Join our server
• GitHub Discussions: Ask questions and share ideas
• Email: admin@cordum.io

Apache-2.0 — see LICENSE.