CDA catches violations at the PR. MCP prevents them before the first line is written. Same governance record. Both ends.
Where each one acts
Catches violations after code is written, at PR time.
Loads design context before the agent writes a line of code.
An agent with only CDA is governed after the fact. An agent with MCP is governed from the start. rkito is the only platform that operationalises both.
Every governed agent session. Every time.
The session protocol runs identically whether the agent is a developer in an IDE or an autonomous pipeline in CI. Same record. Same gates. Same provenance.
Two agent contexts. One governance record.
Bootstrap paths depend on context. IDE agents can use MCP prompts, instruction files, or session hooks. Autonomous agents receive context via orchestrator injection or SDK. Every path leads to the same governed protocol.
IDE Agents
Interactive · Claude Code, Cursor, Copilot Chat, Cline
rkito publishes rkito_session_protocol as a named MCP prompt. When the IDE initialises a session with rkito registered as an MCP server, the protocol is injected into agent context automatically — before the agent takes its first action. Zero developer friction.
Zero friction — when IDE host supports MCP prompts
rkito’s onboarding UI generates a one-time snippet for CLAUDE.md, AGENTS.md, .cursor/rules, or .github/copilot-instructions.md. The developer adds it once; all team members who clone the repo get governed sessions automatically.
At the start of every session, before reading any files,
call get_rkito_protocol() to load the governed interaction
protocol. Then call register_session with your workContextId,
agentInstanceId, and branchRef before proceeding.Works today with all MCP-compatible agents
Claude Code supports hooks that fire before any tool use. rkito provides rkito-mcp-check: a CLI check that verifies session registration before the first tool call fires.
{
"hooks": {
"PreToolUse": [{
"matcher": "*",
"hooks": [{ "type": "command", "command": "rkito-mcp-check --session-registered" }]
}]
}
}Claude Code only — hardest enforcement
Autonomous Agents
Machine · CI pipelines, Devin, orchestrated agents, GitHub Actions
The orchestrator owns bootstrap. It injects RKITO_API_KEY, RKITO_MCP_URL, and a system prompt instruction before the agent starts. Any LLM-based agent that accepts a system prompt. No agent-side feature required.
env:
RKITO_API_KEY: ${{ secrets.RKITO_API_KEY }}
RKITO_MCP_URL: https://mcp.rkito.io/v1Works with any agent that accepts a system prompt
For organisations building custom orchestration pipelines, the rkito Agent SDK wraps session startup, authentication, protocol loading, and context assembly into a single call.
const session = await RkitoSession.start({
apiKey: process.env.RKITO_API_KEY,
workContextId: 'payment-service-update',
branchRef: 'feature/payment-refactor'
})Planned — Enterprise tier
Not documentation. Governed instruction payloads.
A Skill is a versioned, CHI-gated instruction set bound to an Architecture Ledger entity. When an agent retrieves a skill, it receives a conformance bundle — the skill instructions plus the complete set of Lenses applicable to the bound entities. The agent's instruction space is bounded by the same binary criteria that CDA evaluates at PR time.
Authored and consumed within one org. The default. Governance record stays internal.
Discoverable by other orgs in the rkito network. Version-controlled. Forking tracked.
Locally owned derivative. CHI-governed. Provenance link to origin skill preserved.
An agent operating under a skill cannot be instructed to violate a Lens — because the Lens criteria are co-delivered with the instructions.
Everything measurable. Nothing hidden.
Six signals instrument the full governed agent loop — from context access to drift rate to bypass detection. All derived from the same governance record.