aitasks

CLI task management built for AI agents.

aitasks gives AI agents (Claude, Gemini, GPT, etc.) a structured task queue to work from. Agents can claim tasks, log progress notes, verify acceptance criteria with evidence, and hand off work — all via simple shell commands.

Requires Bun ≥ 1.0.0.

Install

bun install -g aitasks

Quick Start

# 1. Initialize in your project
aitasks init

# 2. Create a task
aitasks create --title "Add JWT auth" --desc "Implement user authentication" --priority high --ac "POST /login returns token" --ac "Token expires in 1h"

# 3. Agent picks up work (one-liner with auto-claim)
aitasks next --claim --agent claude-sonnet

# 4. Or claim and start separately
aitasks claim TASK-001 --agent claude-sonnet
aitasks start TASK-001 --agent claude-sonnet

# 5. Log notes along the way
aitasks note TASK-001 "Using bcrypt for hashing — src/auth.ts:L22" --agent claude-sonnet

# 6. Verify acceptance criteria
aitasks check TASK-001 0 --evidence "POST /login returns 200 with token field" --agent claude-sonnet
aitasks check TASK-001 1 --evidence "token exp claim set to now+3600, confirmed in jwt.test.ts:L14"

# 7. Mark done
aitasks done TASK-001 --agent claude-sonnet

Pro tips:

• Use aitasks next --claim --agent <id> to find, claim, and start in one command
• Bulk operations: aitasks done TASK-001 TASK-002 TASK-003 --agent <id>
• Pattern matching: aitasks claim TASK-0* --agent <id> claims all TASK-00x tasks
• Search: aitasks search "auth" finds tasks mentioning authentication
• Undo mistakes: aitasks undo TASK-001
• Delete tasks: aitasks delete TASK-001 (no need to claim first)

Environment Variables

| Variable | Description | |---|---| | AITASKS_AGENT_ID | Default agent ID — avoids passing --agent on every command | | AITASKS_JSON | Set to true to force JSON output globally (useful for scripting) |

export AITASKS_AGENT_ID=claude-sonnet
export AITASKS_JSON=true

Commands

Setup

| Command | Description | |---|---| | aitasks init | Initialize a task database in the current project | | aitasks init --with-review | Initialize with review enforcement (agents cannot mark done without a passing review) | | aitasks onboard | Print or inject agent protocol instructions into CLAUDE.md / AGENTS.md |

Task Discovery

| Command | Description | |---|---| | aitasks list | List all tasks, sorted by priority | | aitasks list --status ready | Filter by status (ready, in_progress, blocked, review, done) | | aitasks next | Show the highest-priority unblocked ready task | | aitasks next --claim --agent <id> | Auto-claim and start the best task | | aitasks show <id> | Full detail on a specific task (includes time tracking) | | aitasks search <query> | Full-text search across titles, descriptions, and notes | | aitasks deps <id> | Show dependency tree (what blocks what) | | aitasks board | Kanban-style board view |

Task Lifecycle

| Command | Description | |---|---| | aitasks create | Create a task (interactive if no flags given) | | aitasks claim <id...> --agent <id> | Claim task(s) - supports patterns like TASK-0* | | aitasks start <id...> --agent <id> | Begin active work on task(s) | | aitasks note <id> <text> | Add an implementation note | | aitasks check <id> <n> --evidence <text> | Verify acceptance criterion n | | aitasks done <id...> --agent <id> | Mark task(s) complete (all criteria must be verified; must be in review status if enforcement is on) | | aitasks review <id...> --agent <id> | Submit task(s) for review (moves to review status) | | aitasks reject <id> --reason <text> | Reject a task in review, send it back to in_progress with feedback | | aitasks unclaim <id> --agent <id> | Release a task back to the pool | | aitasks undo <id> | Undo the last action on a task | | aitasks delete <id...> | Delete task(s) - does not require claiming first |

Note: Commands marked with <id...> support multiple task IDs and pattern matching (e.g., TASK-0*).

Blocking

| Command | Description | |---|---| | aitasks block <id> --on <id,...> | Mark a task as blocked by others | | aitasks unblock <id> --from <id> | Manually remove a blocker |

Agents & History

| Command | Description | |---|---| | aitasks agents | List active agents and their current tasks | | aitasks heartbeat [taskId] | Update agent last-seen timestamp | | aitasks log <id> | Full event history for a task | | aitasks export --format json | Export all tasks as JSON | | aitasks export --format csv | Export all tasks as CSV |

Database

| Command | Description | |---|---| | aitasks db status | Show database health and stats |

create Flags

aitasks create \
  --title "My task" \
  --desc "Longer description" \
  --ac "Returns 200" \          # Acceptance criterion (repeatable, at least one required)
  --priority high \             # critical | high | medium | low
  --type feature \              # feature | bug | chore | spike
  --parent TASK-001 \           # Parent task ID (optional)
  --agent $AITASKS_AGENT_ID     # Agent creating the task (logged in event history)

Agent Protocol

When you run aitasks init, it automatically injects a full agent protocol into CLAUDE.md, AGENTS.md, or GEMINI.md (whichever exists, or creates AGENTS.md). This tells the AI agent exactly how to use aitasks.

You can also inject/view it manually:

aitasks onboard               # print to stdout (reflects current review_required setting)
aitasks onboard --append      # append to detected agent file
aitasks onboard --file MY.md  # append to a specific file
aitasks onboard --json        # output as JSON string

The injected instructions automatically adapt to the project's review enforcement setting — if --with-review is enabled, agents receive the full review workflow (with aitasks review, sub-agent approval, and aitasks reject) instead of the standard completion flow.

Review Enforcement

Enable a mandatory review gate so agents cannot mark tasks done without an explicit approval step:

# Enable at init time
aitasks init --with-review

# Enable on an existing project (also replaces the agent instructions file with review-aware version)
aitasks init --with-review  # safe to re-run; updates DB setting and rewrites agent file

A task is only complete when its status is done. No other status — verified criteria, review, in_progress — counts as complete. Agents must reach done to consider a task finished.

How it works:

1. Agent completes work and verifies all acceptance criteria with evidence
2. Agent submits for review — task moves to review status:

   aitasks review TASK-001 --agent $AITASKS_AGENT_ID

   The command output explicitly instructs the agent to immediately spawn a review sub-agent. The task is not complete at this point.
3. The agent immediately spawns a review sub-agent to inspect the implementation:
   • Approves — moves task to done:

     aitasks done TASK-001 --agent review-agent

   • Rejects — sends it back to in_progress with feedback:

     aitasks reject TASK-001 --reason "Missing error handling for 404 case" --agent review-agent

4. If rejected, the original agent addresses the feedback, re-verifies criteria, and repeats from step 2.

Enforcement: aitasks done and aitasks update --status done both block if the task isn't already in review status. The gate cannot be bypassed.

Board: Tasks in review status appear in the IN PROGRESS section with a ◈ magenta indicator so they're visually distinct from actively-worked tasks.

Data Storage

aitasks init creates a .aitasks/ directory in your project root containing a SQLite database. Add .aitasks/ to your .gitignore if you don't want to commit task data, or commit it to share tasks across your team.

License

MIT