🦞 clawdo — The missing todo list for AI agents

Your agent has memory files, cron jobs, and chat. It has no todo list.

No way to say "do this when you get to it." Not "do this at 14:00 UTC." Not "do this right now in this conversation." Just... remember to do it. Track it. Pick it up when there's a gap.

That's clawdo.

Install

clawhub install clawdo    # installs skill + docs into your workspace
npm install -g clawdo     # install the CLI binary

Requirements: Node.js ≥18

Quick Start

# Capture a task
clawdo add "update dependencies" --urgency soon

# Agent checks its queue (heartbeat, cron, conversation — wherever)
clawdo inbox --format json

# Agent works it
clawdo start a3f2
clawdo done a3f2 --json

add → inbox → start → done. Persistent state in SQLite. Every command has --json so agents parse structured output, not terminal art.

Where it fits

clawdo works everywhere agents work:

• Heartbeat loops — "anything in my queue? let me do it between checks"
• Cron jobs — "every hour, process one task"
• Conversations — "J mentioned fixing the auth module, let me capture that"
• Pipes and sub-agents — non-TTY safe, no interactive prompts

Heartbeat integration example

# In HEARTBEAT.md — runs every ~30 minutes
TASKS=$(clawdo inbox --format json)
AUTO=$(echo "$TASKS" | jq '.autoReady | length')

if [ "$AUTO" -gt 0 ]; then
  TASK=$(clawdo next --auto --json | jq -r '.task.id')
  clawdo start "$TASK" --json
  # ... do the work ...
  clawdo done "$TASK" --json
fi

Autonomy levels

Tasks can be tagged with permission tiers that control what the agent is allowed to do unsupervised:

Default: collab (safe).

Key rule: Autonomy is a permission, not a suggestion. Once set, agents can't change it. If an agent fails 3 times, autonomy demotes to collab. Safety only moves down, never up.

Agents propose, humans approve. Agent tasks always start as proposed. The human runs clawdo confirm \x3Cid> or it doesn't happen.

Usage

For humans

# Add tasks — inline metadata parsing
clawdo add "deploy new API +backend auto-notify now"
#           └── text ──────┘ └project┘ └─level──┘ └urg┘

# View
clawdo list                       # active tasks
clawdo list --status proposed     # agent suggestions
clawdo next                       # highest priority

# Review agent proposals
clawdo confirm \x3Cid>               # approve
clawdo reject \x3Cid>                # reject

# Work
clawdo start \x3Cid>
clawdo done \x3Cid>
clawdo done abc,def,ghi           # complete several

For agents

# Check inbox (structured)
clawdo inbox --format json

# Propose work
clawdo propose "add input validation" --level auto --json

# Execute
TASK=$(clawdo next --auto --json | jq -r '.task.id // empty')
if [ -n "$TASK" ]; then
  clawdo start "$TASK" --json
  # ... do the work ...
  clawdo done "$TASK" --json
fi

The inbox returns: autoReady, autoNotifyReady, urgent, overdue, proposed, stale, blocked.

Inline syntax

clawdo add "fix auth bug +backend @code auto soon"

• +word → project
• @word → context
• auto / auto-notify / collab → autonomy level
• now / soon / whenever / someday → urgency
• due:YYYY-MM-DD → due date

Security

• Immutable autonomy — agents cannot escalate permissions
• Proposal limits — max 5 active, 60s cooldown
• Prompt injection defense — input sanitization, parameterized SQL
• Audit trail — append-only log of every state change
• Secure IDs — crypto.randomInt(), no modulo bias

Resources

• GitHub: https://github.com/LePetitPince/clawdo
• npm: https://www.npmjs.com/package/clawdo
• Full docs: clawdo --help

License

MIT