Let AI agents manage real work without letting them run wild.
pm.tasks is an agent-safe, MCP-native task tracker where humans and AI agents share one source of truth — with strict workflow rules, tiny tool responses, and auditable project history.
$ update_task({ id: 14, status: "done" }) ✕ rejected todo → done is not allowed legal next: in_progress · cancelled
$ update_task({ id: 14, status: "in_progress" }) ✓ allowed todo → in_progress · started_at set
$ add_dependency({ task_id: 15, depends_on_id: 14 }) ✕ rejected cycle detected: #15 → #14 → #15
$ update_task({ id: 14, status: "to_review" }) ✓ allowed in_progress → to_review · awaiting human
AI agents break normal project-management tools in three ways.
Traditional work-tracking tools were designed for humans clicking dashboards. Agents need something smaller, stricter, and easier to reason about.
They burn context on noise
Huge APIs, custom fields, metadata, labels, boards, sprints, and histories get pushed into agent calls. The token cost adds up fast.
They operate through human-shaped workflows
Most PM tools assume someone is clicking, filtering, scanning, and dragging. Agents need small deterministic actions, not UI archaeology.
They drift from team reality
Agent notes spread across chats, markdown files, GitHub comments, and private memory. Humans lose visibility, and trust disappears.
Before: scattered agent state
The first demos feel magical. Then project state starts leaking everywhere.
After: one controlled work layer
Agents create, update, comment, and progress work through rules humans can inspect.
Agents need rails, not dashboards.
AI agents do not need drag-and-drop boards, emoji reactions, or twelve custom fields before they can close a bug. They need a small set of safe actions and a workflow your team can trust.
Build an agent-safe workflowEverything an agent needs. Nothing it doesn't.
Built small on purpose, because every extra field is another chance for cost, confusion, or agent error.
Strict status transitions
Agents cannot skip review, close the wrong task, or invent unsupported states. The workflow is enforced by the tool, not by prompt discipline.
Shared human-agent state
Humans and agents work from the same tasks, comments, statuses, and history — no hidden scratchpads or private project state.
Token-friendly responses
Small tool payloads reduce AI operating cost and help agents focus on the work instead of irrelevant project noise.
MCP-native
Connect any MCP-compatible client directly. No scraping, no fragile UI automation, and no giant project-management API.
Enforced dependencies
Tasks declare what must happen first, and circular work is rejected by the binary — agents cannot run ahead of the plan.
Definition of done
Every task carries explicit acceptance criteria, so agents know exactly what "done" means — and humans review against it.
Responsible agent workflows need visible project state.
pm.tasks helps companies move from AI experiments to AI execution without losing control, auditability, or engineering trust.
Lower AI operating cost
Smaller responses mean fewer tokens spent every time an agent checks, updates, or summarizes work.
Better auditability
Know what agents touched, when they touched it, and how the work moved through the workflow.
Safer automation
Guardrails stop agents from making unauthorized workflow jumps or corrupting project status.
Faster AI adoption
Give agents a focused execution layer without forcing your team to replace its entire PM system.
Built for agent work, not board work.
pm.tasks is not trying to be a full enterprise PM suite. It is the controlled work layer your AI agents can safely use.
| Capability | Traditional PM tools | pm.tasks |
|---|---|---|
| Primary user | Humans clicking dashboards | Humans + AI agents sharing work state |
| Agent access | Large APIs, plugins, scraping, or browser automation | MCP-native from day one |
| Context size | Often bloated with fields and metadata | Small, token-conscious responses |
| Workflow control | Flexible but inconsistent | Strict enforced status transitions |
| Agent reliability | Depends heavily on prompts | Rules enforced by the product |
| Best use | Full project management for human teams | AI-agent work execution and coordination |
From agent chaos to controlled execution.
A simple workflow for teams that want AI agents to track real work safely.
Join early access
Tell us what AI-agent workflow you are trying to manage.
Connect your MCP client
Point your agent at a small, deterministic issue-tracking surface.
Let agents create and update tasks
Agents pick up tasks with a clear definition of done, comment as they go, and progress status safely.
Humans review and approve
Keep human oversight in the loop with shared task history and enforced transitions.
Designed for the messy middle of human-agent work.
Use pm.tasks wherever AI agents create, triage, update, or hand off project work.
Autonomous maintenance tasks
Track dependency upgrades, refactors, and cleanup work done by agents.
AI coding-agent bug triage
Let coding agents file issues and spawn fix tasks while humans retain review control.
QA agents reporting failures
Test agents can file structured issues with severity, reproduction notes, and status.
Support-to-engineering automation
Convert customer problems into controlled engineering tasks without losing context.
Multi-agent coordination
Give multiple agents one shared work state instead of scattered private memory.
AI governance pilots
Give business and engineering leaders visibility into how agents move work forward.
Who this is for
- Teams testing AI agents in engineering workflows
- Developers using MCP-compatible clients
- Companies worried about agent-created project chaos
- AI automation builders who need clean work state
- Engineering teams that want auditable agent work
Who this is not for
- Teams wanting a full Jira replacement today
- Companies with only human-only project boards
- Workflows requiring highly customized enterprise PM fields from day one
- Teams not yet experimenting with agentic work
- Organizations that prefer prompt-only workflow discipline
Build your first agent-safe task workflow.
Join early access if you are experimenting with AI agents, MCP, or human-agent engineering workflows.
Using an AI agent? Let it sign you up instead.
You're on the list!
We'll email you when early access opens. No spam, no lists, no cookies.