Skip to content

OpenClaw Integration

OpenClaw is strong when an agent can act across the tools and channels you already use. DevCD matters one step earlier: it stops a fresh OpenClaw session from starting cold.

The intended moment is Turn-0 continuity. Before a new OpenClaw agent asks what the goal is, what failed last, or what to try next, it should read DevCD's local, policy-filtered Action Packet and start from there.

DevCD integrates with OpenClaw through the standard local MCP stdio boundary. It is not an OpenClaw plugin, does not mutate OpenClaw configuration, and does not require write-capable MCP tools.

Why OpenClaw Users Care

  • OpenClaw already gives agents reach. DevCD gives them continuity.
  • OpenClaw can act across channels and tools. DevCD tells the next session what not to rediscover.
  • OpenClaw is strongest when the first turn starts from the current goal, latest failure, blocker, and next action instead of a recap request.

The desired outcome is simple: OpenClaw starts warm.

Fastest Local Proof

If you want to prove the DevCD side first, start with the checked-in Action Packet demo:

bash devcd agentic action-packet-demo --events examples/agentic-action-packet/sample-events.jsonl

Then verify the OpenClaw-facing MCP shape locally:

bash devcd integrations openclaw --smoke-test

That sequence proves the handoff surface first and the MCP wiring second.

Verified DevCD Side

The DevCD side can be checked without installing OpenClaw:

bash devcd integrations openclaw --smoke-test

This command prints a copyable local MCP snippet and verifies the shape of the read-only MCP server in-process. It checks initialize, resources, tools, and prompts without editing ~/.openclaw/openclaw.json.

The MCP command OpenClaw should launch is:

bash devcd mcp serve

DevCD currently exposes read-only resources such as Continuity Packet, Action Packet, Session Contract, Agent Handoff Packet, work state, policy summary, recent timeline, and withheld-context summaries. It exposes no write-capable MCP tools.

For OpenClaw users, the important point is not the full resource list. The important point is that Turn-0 can begin from Action Packet and Session Contract, while the deeper continuity surfaces stay available only when needed.

Configuration Shape

Use the CLI to print the current snippet:

bash devcd integrations openclaw devcd integrations openclaw --json

The relevant shape is:

json { "mcp": { "servers": { "devcd": { "command": "devcd", "args": ["mcp", "serve"] } } } }

Add this manually to OpenClaw configuration only when you choose to connect the local runtime.

This keeps the trust boundary explicit: DevCD prints the snippet, you decide whether OpenClaw should read it.

What An OpenClaw Agent Should Read

At the start of a session, the agent should prefer:

  1. devcd://context/action-packet for next-action guidance.
  2. devcd://context/session-contract for the next action, verification command, clean-state requirement, and context-reference loading hints.
  3. devcd://context/policy-summary before asking for withheld context.
  4. devcd://context/continuity-packet for domain-neutral continuity if the Action Packet is not ready.
  5. devcd://context/withheld-context to understand what was intentionally not included.

If an older workflow still expects it, devcd://context/agent-handoff-packet is kept as a legacy developer handoff contract.

OpenClaw Skill Draft

The first OpenClaw-native artifact is the small Skill draft at skills/devcd-continuity/SKILL.md. Its job is not to install DevCD or change OpenClaw config. Its job is to teach the agent the first move:

  1. Read devcd://context/action-packet.
  2. Read devcd://context/policy-summary.
  3. Start with the goal, next action, blockers, do_not_repeat, and policy note.
  4. Ask only for confirmation or correction.

This is the intended OpenClaw moment: the agent begins warm instead of turning the user into the continuity layer.

Current Limitation

The DevCD MCP server and config snippet are locally verified. The full OpenClaw-gateway end-to-end flow is intentionally not claimed here until a real OpenClaw runtime has loaded and read the DevCD resources.

Until that verification exists, DevCD should not be described as:

  • an official OpenClaw plugin;
  • a ClawHub package;
  • a ClawHub skill;
  • an OpenClaw-native runtime component;
  • an integration that automatically edits OpenClaw config.

Artifacts

  • OpenClaw Skill draft: skills/devcd-continuity/SKILL.md
  • MCP proof notes: examples/openclaw-mcp-context/README.md
  • Packaging analysis: docs/devcd/openclaw-packaging-spike.md
  • Draft use case: docs/devcd/openclaw-usecase-draft.md
  • Release readiness: docs/devcd/release-readiness.md