claw-doing

See what your agent is doing — right now.

claw-doing is an OpenClaw plugin that gives long-running agent tasks real-time progress visibility. Instead of waiting silently for minutes while your agent works, you get a live step-by-step progress feed — right in your messaging channel.

The problem

When you ask OpenClaw to do something complex — research a topic, process files, build a multi-step pipeline — the interaction goes silent:

You:   "Research AI agent frameworks and write a report"
Agent: [silent for 8 minutes]
You:   "...is it still running?"
Agent: [finally replies, or times out]

You have no idea what's happening, whether it's stuck, or when it will finish.

What claw-doing does

With claw-doing installed, the same interaction looks like this:

You:   "Research AI agent frameworks and write a report"

Agent: Research AI agent frameworks
       Task ID: `a3f8c2d1-...`

       Plan — 4 steps:
         1. Search and collect sources (~3 min)
         2. Extract key findings (~2 min)
         3. Cross-reference and validate (~2 min)
         4. Write and deliver report (~3 min)

       [░░░░░░░░░░] 0/4
       Starting now...

       ──── (3 minutes later) ────

       [███░░░░░░░] 1/4 — elapsed: 3m 12s

       ✓ Step 1: Found 47 sources, filtered to 14 high-quality references
       ⚠ Note: Two sources were paywalled — used abstracts only

       Next: Step 2 — Extract key findings

       ──── (2 minutes later) ────

       [██████░░░░] 2/4 — elapsed: 5m 8s

       ✓ Step 2: Extracted 31 key data points across 8 frameworks

       ... and so on ...

       [██████████] 4/4 — 11m 42s total

       ✅ Completed successfully

       Compiled a 2,800-word report covering 8 frameworks with 12 citations.
       Key finding: LangGraph leads on multi-agent orchestration; OpenClaw leads on channel integrations.

       Deliverables:
         • memory/2026-03-29-ai-agent-frameworks.md

Installation

openclaw plugins install claw-doing

OpenClaw checks ClawHub first, then falls back to npm automatically.

After installation, restart your OpenClaw gateway to load the plugin.

Enable the bootstrap instructions (recommended)

The plugin ships with bootstrap/TASK_TRACKER.md — a set of instructions that teaches your agent when and how to use the tracking tools. Add it to your OpenClaw workspace bootstrap:

{
  "hooks": {
    "internal": {
      "enabled": true,
      "entries": {
        "bootstrap-extra-files": {
          "enabled": true,
          "paths": ["node_modules/claw-doing/bootstrap/TASK_TRACKER.md"]
        }
      }
    }
  }
}

Or copy it manually to your workspace — locate TASK_TRACKER.md inside the installed package under node_modules/claw-doing/bootstrap/ and place it in your OpenClaw workspace directory (typically ~/.openclaw/workspace/).

Then add TASK_TRACKER.md to your bootstrap file list so the agent loads it on start.

How it works

claw-doing registers three agent tools. The agent calls them as it works — no separate dashboard or server required.

Tools

task_plan

Called at the start of a multi-step task. Registers the plan and returns a task ID.

| Parameter | Type | Description | |-----------|------|-------------| | title | string | Short title for the task | | steps | Step[] | Ordered list of steps | | steps[].index | number | Step number (1-based) | | steps[].description | string | What this step does | | steps[].estimatedMinutes | number? | Optional time estimate |

Returns: task ID + formatted plan for the user.

task_step_done

Called after each step completes. Updates progress and notifies the user.

| Parameter | Type | Description | |-----------|------|-------------| | taskId | string | Task ID from task_plan | | stepIndex | number | Which step was just completed | | summary | string | What was accomplished | | deviation | string? | Unexpected findings (optional) |

Returns: progress bar + step summary.

task_complete

Called when the entire task finishes (or fails). Delivers the final report.

| Parameter | Type | Description | |-----------|------|-------------| | taskId | string | Task ID from task_plan | | outcome | "success" \| "partial" \| "failed" | Task outcome | | summary | string | User-facing completion summary | | artifacts | string[]? | Output files or results produced |

Returns: final progress bar + summary + deliverables list.

Task state

Each task is persisted as a JSON file at:

~/.openclaw/workspace/claw-doing/tasks/<taskId>.json

This means task state survives agent restarts and can be inspected or archived.

Configuration

Optional config in your openclaw.config.json:

{
  "plugins": {
    "claw-doing": {
      "storageDir": "/custom/path/to/tasks",
      "progressBarWidth": 12
    }
  }
}

| Option | Default | Description | |--------|---------|-------------| | storageDir | ~/.openclaw/workspace/claw-doing/tasks/ | Where task JSON files are stored | | progressBarWidth | 10 | Width of the █░ progress bar |

You can also set the storage directory via environment variable:

CLAW_DOING_STORAGE_DIR=/my/custom/path openclaw start

Design principles

Zero deployment. Unlike dashboard-based solutions, claw-doing requires no external server, database, or Docker container. Install the plugin and it works.

In-channel delivery. Progress updates arrive in the same messaging channel you're already using — Discord, Telegram, WhatsApp, web chat. You don't need to open a separate UI.

Agent-driven, not hook-driven. The agent explicitly calls task_step_done after each step. This makes progress reporting intentional and semantically rich (the agent decides what counts as "done" and what to say about it), rather than relying on low-level hook instrumentation that may fire at the wrong granularity.

Persistent state. Task records are written to disk as JSON. They can be read, archived, or used by other tooling (dashboards, scripts) independently.

Comparison with similar projects

| | claw-doing | ClawBeacon | |---|---|---| | Form factor | OpenClaw plugin (zero deployment) | Separate web app (Docker/Railway) | | Where you see progress | Your chat channel | A Kanban dashboard in a browser | | Setup | openclaw plugins install claw-doing | Clone, deploy backend + frontend | | Best for | Individual users, any task | Teams managing multiple agents |

Development

git clone https://github.com/seekcontext/claw-doing
cd claw-doing
npm install
npm run dev        # watch mode
npm run typecheck  # type check without emitting
npm run build      # compile to dist/

Project structure

claw-doing/
├── index.ts                      # Plugin entry — registers all tools
├── src/
│   ├── types.ts                  # Shared Task and TaskStep interfaces
│   ├── store/
│   │   └── task-store.ts         # File-based task state (create, update, complete)
│   ├── tools/
│   │   ├── task-plan.ts          # task_plan tool handler
│   │   ├── task-step-done.ts     # task_step_done tool handler
│   │   └── task-complete.ts      # task_complete tool handler
│   └── utils/
│       └── format.ts             # Progress bar, elapsed time, outcome labels
├── bootstrap/
│   └── TASK_TRACKER.md           # Agent behavior instructions
├── openclaw.plugin.json          # Plugin manifest
└── openclaw.d.ts                 # Type stubs for openclaw peer dependency

Roadmap

• [ ] before_tool_call hook to auto-detect sessions_spawn and link sub-agents to parent tasks
• [ ] Deviation detection: flag steps where output diverges significantly from expected
• [ ] task_stuck tool: pause execution and ask the user for guidance mid-task
• [ ] Task history CLI: openclaw doing list to review past task records
• [ ] Integration with OpenClawTaskStatusV1 (openclaw/openclaw#42385) when merged

Contributing

Pull requests welcome. Please open an issue first for significant changes. Repository: github.com/seekcontext/claw-doing

License

MIT