Skip to content

OpenClaw Agent Skill Integration

Author: PlanExe Team
Date: 2026-02-11
Status: Proposal
Audience: OpenClaw Developers, Agent Architects

Pitch

Package PlanExe as a standardized OpenClaw skill that turns agents into project managers: generate plans in the cloud, execute locally on edge devices, and report progress back into PlanExe.

Why

Edge agents have sensors and actuators but low compute. Cloud agents can plan but lack physical access. A unified skill bridges this split and enables coordinated execution.

Problem

  • Edge agents lack LLM capacity to generate robust plans.
  • Cloud agents cannot directly execute physical tasks.
  • There is no consistent interface for plan generation and task execution.

Proposed Solution

Create a $PlanExeSkill for OpenClaw that:

  1. Drafts plans via PlanExe Cloud.
  2. Breaks plans into executable tasks.
  3. Routes tasks to edge or human executors.
  4. Reports results and updates the plan state.

Architecture

OpenClaw Agent
  -> PlanExe Skill
     -> MCP Client
        -> PlanExe Cloud
           -> Plan JSON
     -> Task Executor
  -> Result Reporter

Skill Manifest (skill.json)

{
  "name": "PlanExe Project Manager",
  "version": "1.0.0",
  "description": "Gives the agent the ability to plan, budget, and track complex projects via the PlanExe Cloud.",
  "permissions": ["network_access"],
  "mcp_tools": [
    "prompt_examples",
    "task_create",
    "task_status",
    "task_stop",
    "task_file_info"
  ]
}

Skill Capabilities (MCP Tools)

The PlanExe MCP exposes the following real tools via https://mcp.planexe.org/mcp:

prompt_examples() — Get example prompts

  • Input: none
  • Output: List of example planning prompts
  • Used in: Skill initialization and user guidance

task_create(prompt, speed_vs_detail, model_profile, user_api_key?) — Start a planning task

  • Input:
  • prompt (string): Natural language planning request
  • speed_vs_detail (enum): "ping" | "fast" | "all"
  • model_profile (enum): "baseline" | "premium" | "frontier" | "custom"
  • user_api_key (optional string): Override environment API key
  • Output: task_id for polling and result retrieval
  • Used in: Step 1 of planning workflow

task_status(task_id) — Poll planning progress

  • Input: task_id (string)
  • Output: status (queued|running|completed|failed), progress %, estimated time remaining
  • Used in: Polling loop every 5+ minutes (plans take 15-20+ min)

task_stop(task_id) — Cancel a running task

  • Input: task_id (string)
  • Output: Confirmation of cancellation
  • Used in: Early termination / user interruption
  • Input:
  • task_id (string)
  • artifact (enum): "report" | "zip"
  • Output: download_url
  • Used in: Retrieving completed plans after task finishes

Agent-to-Agent Protocol

  • EdgeBot detects a local issue (low water).
  • EdgeBot requests a plan from CloudBot.
  • CloudBot calls task_create() via PlanExe MCP and receives task_id.
  • CloudBot polls task_status(task_id) every 5+ minutes.
  • CloudBot calls task_file_info(task_id, "report") to get download URL when complete.
  • EdgeBot executes local steps from the plan and reports back.
  • CloudBot can call task_stop(task_id) if execution is cancelled.

Integration Points

  • Uses PlanExe MCP interface for plan creation.
  • Feeds execution data to assumption drift and readiness scoring.
  • Works with distributed physical task dispatch protocol.

Success Metrics

  • Installation rate: % of OpenClaw instances with PlanExe skill.
  • Plan completion rate without human intervention.
  • Mean time from goal to first actionable task.

Risks

  • Over-reliance on cloud connectivity.
  • Misaligned task interfaces between cloud and edge.
  • Skill misuse without governance or budget limits.

Future Enhancements

  • Offline plan caching for intermittent connectivity.
  • Capability-aware task routing across multiple agents.
  • Automatic escalation to humans for high-risk tasks.

Detailed Implementation Plan

Phase A — Skill Packaging and Contracts (1–2 weeks)

  1. Define skill manifest and MCP tool bindings:
  2. prompt_examples — Provide example planning prompts
  3. task_create — Initiate planning task with goal and parameters
  4. task_status — Poll task progress (required for async workflows)
  5. task_stop — Cancel long-running tasks

  6. task_file_info — Retrieve generated plan artifacts

  7. Add JSON schema validation for tool inputs/outputs.

  8. Version the skill API separately from PlanExe core API.

Phase B — MCP Bridge and Runtime Adapter (2–3 weeks)

  1. Implement MCP client wrapper inside skill runtime.
  2. Add retry/backoff and circuit-breaker handling for cloud calls.
  3. Normalize errors into agent-friendly remediation messages.

Phase C — Execution Loop and State Sync (2 weeks)

  1. Build local task executor abstraction for edge environments.
  2. Add plan-state sync protocol:
  3. pull next action
  4. execute/report

  5. reconcile conflicts

  6. Support partial completion and blocked states.

Phase D — Safety, Budgets, and Governance (2 weeks)

  1. Add execution policy profiles by risk tier.
  2. Add per-run budget controls and token/cost caps.
  3. Add human escalation hooks for high-impact tasks.

Integration points

  • OpenClaw skill registry
  • PlanExe MCP cloud endpoints
  • Execution readiness and evidence ledger modules

Operational safeguards

  • Offline queue with replay upon reconnect
  • Idempotency keys for duplicate submissions
  • Signed task receipts for non-repudiation

Validation checklist

  • End-to-end tool contract tests
  • Network partition resilience tests
  • Duplicate event/retry safety tests
  • Human escalation path latency and success tests