The Cordum Agent Protocol (CAP).
MCP governs what agents say. CAP governs what agents do.
CAP is an open, language-agnostic wire protocol for AI agent governance. It standardizes how jobs are submitted, how heartbeats are reported, and how policies are enforced.
Unified governance requires a standard contract.
Every agent framework invents its own wire format. This makes it impossible to enforce global policy or maintain a consistent audit trail across heterogeneous agents.
Protocol-First Design
CAP standardizes the BusPacket envelope, typed payloads, and opaque memory pointers. This ensures that a scheduler can evaluate a job from a Python agent and dispatch it to a Go worker without custom translation layers.
Safety as a Primitive
Unlike generic message buses, CAP includes first-class support for policy evaluation hooks, human-in-the-loop approvals, and runtime constraints.
message BusPacket {
string trace_id = 1;
string sender_id = 2;
google.protobuf.Timestamp created_at = 3;
oneof payload {
JobRequest job_request = 10;
JobResult job_result = 11;
Heartbeat heartbeat = 12;
JobProgress job_progress = 13;
}
}Adopt CAP in your stack.
The Go runtime SDK is production-ready. A Python guard SDK is available in beta. The wire format is language-agnostic, so any language that speaks NATS can participate.
Go
stablePrimary Runtime
import "github.com/cordum/cordum/sdk/runtime"
w := runtime.NewWorker(runtime.Config{
Topic: "job.incident.enrich",
})
w.Handle(func(ctx *runtime.Context) error {
// Business logic here
return ctx.Succeed(result)
})Python
betaGuard SDK
from cordum_guard import Guard
guard = Guard(topic="job.analyst.run")
@guard.handle()
def on_job(ctx):
# Process with your framework
ctx.succeed({"status": "complete"})What CAP covers.
The protocol spans three tiers of capability, from basic job dispatch through production-grade governance.
The minimum set of primitives for agent-to-platform communication.
- Job lifecycle (submit / result)
- Worker heartbeats
- BusPacket envelope
- Trace propagation
Advanced orchestration and human-in-the-loop patterns.
- Workflow DAGs
- Approval gates
- Checkpoint heartbeats
- Saga compensation
Deep integration for production platforms and runtime safety.
- Context engine (memory)
- Artifact pointers
- Schema registry
- Packet signing & verification
MCP governs what agents say. CAP governs what agents do.
Model Context Protocol (MCP) and Cordum Agent Protocol (CAP) are complementary standards for the agentic era.
Standardizes tool discovery and invocation within a model session. It determines how an agent asks to use a tool.
Standardizes job lifecycle, policy, and audit across the platform. It determines if an agent is allowed to execute an action.
Adopt the standard.
CAP is source-available and append-only. Help us build the governance layer for autonomous AI agents.