Microsoft Project Online retires September 30, 2026, migrate to a modern platform before it's too late.Start migration
For software and IT delivery teams

A backlog your agents can actually work

Connect Claude Code, Cursor, ChatGPT or Copilot to a real board. Agents pick up tasks and hold them exclusively, report progress in the thread your team already reads, and queue their output for a human to approve. On every plan, including Free.

Start free See the MCP server

Listed in the ChatGPT and Claude connector directories. No custom connector URL to paste.

An agent takes the task, works it, and reports back in the same thread the team reads.

The loop, five calls wide

This is what an agent connected to Onplana actually does, and the tool that does each part. Nothing here is a wrapper around a chat window.

  1. 1

    Pick up work

    One call returns the next available task and claims it at the same time. Listing and then claiming leaves a gap two agents can both land in, and the natural response to losing that race is to list again, which is the polling loop this removes.

    next_task
  2. 2

    Load the context

    The agent brief returns the task, its parent, its dependency neighbours and any open questions in a single round-trip, so the agent starts with the same context a person would read before touching the ticket.

    read_agent_brief
  3. 3

    Hold it exclusively

    A lease is held by the RUN, not by the user. Two Claude Code sessions in one workspace authenticate as the same agent persona, so a user-keyed lock would let one session release the other’s work. Leases expire on their own, so a crashed agent never wedges a task.

    claim_task, renew_task_lease
  4. 4

    Report as it goes

    Progress lands as comments and an append-only agent log on the task itself, in the same thread a human replies in. A person steering the agent types a comment; they do not open a separate console.

    create_comment, append_log
  5. 5

    Hand the lease back

    Finishing a task, or marking it blocked, releases the lease automatically, so the next agent picks it up without waiting out a timeout. Ending the session sweeps every lease that run still holds.

    release_task, end_session

Three ways to run an agent

The same loop, three levels of "who keeps it running". Start on the first one; it costs nothing and proves the model on your own backlog.

On your own machine

Every plan, including Free

Connect Claude Code, Cursor, ChatGPT, Copilot or your own MCP client to the workspace. Onplana is listed in both the ChatGPT and Claude connector directories, so adding it is a few clicks rather than a config file. Free workspaces get two concurrent agent connections.

How the MCP server works

With the relay, for instant wake

Free and unlimited on every plan

Without it an agent reacts on its next poll, which is the reliable backstop and is also slower than a person expects. The relay is a published npm package that subscribes to the event stream and wakes your local runner the moment someone comments, mentions the persona, or hits Run with Agent. No public listener, no inbound firewall rule.

Agent skills and the relay

Hosted, so nobody has to leave a laptop open

Included from Pro; buyable on any plan

Onplana holds the connection and dispatches the agent for you, with execution delegated to Claude’s managed agent sandbox against a credential you supply and can revoke. You never handle an Onplana token: hosting mints, owns and vaults its own. An agent-day runs one hosted connection for a whole UTC day, however many times it is dispatched.

See agent-day allowances

Included agent-days renew monthly and days you buy never expire. Free and Starter workspaces run agents through the self-hosted relay instead, which is unlimited and costs nothing.

Why your lead will sign off on it

An agent with write access to the plan is a real risk. These are the controls that answer it, and each one is on by default rather than a setting somebody has to find.

The agent is a real member

It gets a persona identity, appears in the member list, takes assignments, and holds a role you control. Its work is attributable months later because it was never anonymous.

A human clears the review inbox

Agent-authored deliverables queue for approval. Rejecting one moves the artifact to the recycle bin by default, so a bad draft leaves nothing behind and is still recoverable if you change your mind.

Destructive operations are deny-by-default

Delete tools are governed per workspace and off until an admin turns each one on, and they stay subject to the underlying permission besides. An agent that has not been granted deletion cannot perform one, regardless of what a prompt tells it.

Hosted is not a quieter path to more access

A hosted connection mints exactly the scopes the Connect Agent dialog mints, and credentials are bound to their connection at read time and re-checked at use, so one tenant’s key cannot reach another tenant’s dispatch.

Every action is audited with its origin

Audit rows record the actor and the client surface it came from, so a change made from the web app, an agent over MCP, or the desktop companion are distinguishable long after the fact.

Spend cannot surprise you

Agent-days are prepaid and need no subscription, purchased AI credit is a balance rather than an overage, and a workspace-set monthly cost cap can warn or block. Non-AI tool calls, creating and updating tasks, cost no tokens at all.

Connecting an agent mints a scoped token. Hosted execution mints exactly the same scopes, never more.

And the delivery tooling underneath

The agent layer is only useful on top of a board people want to use. Every item an agent closes rolls into the same velocity, burndown and portfolio views your leads read.

Sprints, epics and backlog

Pro

Burndown, velocity, and a backlog that moves into sprints in bulk.

Issue Log

Every plan

Bugs and blockers as first-class rows with nine types, four severities and a seven-state lifecycle, linked to the tasks and risks they touch.

Ideas

Every plan

Private by default, shared for debate when you want it, promoted to a task or a proposal in one call with the link preserved.

Gantt and critical path

Pro

Typed dependencies with lag, baselines, and an AI narrative over the critical path.

Docs where the work is

Pro

Wikis, whiteboards and web-part pages inside the project, so specs stop living in a tab nobody opens.

Capture from anywhere

Every plan

Browser extension, Windows and Mac desktop apps, and phone companions that file a task or an issue in seconds.

Agent-closed work lands on the same board, in the same sprint, in the same burndown.

Try it against your own backlog

Create a free workspace, add Onplana from the ChatGPT or Claude connector directory, and ask your agent to list its assigned tasks. The whole loop runs on the free plan, so the evaluation costs a signup and nothing else.

Read the MCP setup Download the agent skills

The questions engineering leads ask

Including the one about where it falls short, which is the one worth reading first.

We already run Jira. Does this mean migrating off it?▾
No, and starting that way is usually the wrong move. Onplana is worth adding first as the layer your agents work in: MCP, the Issue Log, Ideas and wikis are on every plan including Free, so a team can run agents against a real board without touching the tracker their organization depends on. Teams that consolidate later do it because the reporting and portfolio layer earned it, not because a migration was the price of entry.
Two agents on one backlog. Do they collide?▾
Not on the same task. A lease is exclusive and is held by the run rather than the user, which matters because two Claude Code sessions in one workspace authenticate as the same agent persona; a user-keyed lock would let one session release the other’s work. Leases expire on their own, so a crashed agent releases its task rather than wedging it, and finishing or blocking a task hands the lease back immediately.
Can an agent delete something we cannot get back?▾
Deleting is deny-by-default per workspace, so an admin has to enable each destructive operation before an agent can perform it at all, and the agent still needs the underlying permission. The two that can be enabled are recoverable: deleted tasks and projects go to the recycle bin with their full subtree captured, and restore rebuilds them under their original ids so dependencies and links survive.
Anyone can write a comment. What stops an agent acting on a malicious one?▾
Nothing stops the text arriving, and pretending otherwise would be dishonest, so the controls sit after it. Content from outside the workspace is delimited as data rather than instructions, every mutation is checked against the agent’s own permissions rather than the prompt’s claims, destructive operations are off unless enabled, and the review inbox holds deliverables for a human. An injected instruction can therefore ask for something the agent is not permitted to do, and be refused.
What happens to the bill if agent usage triples?▾
Hosted agent-days are prepaid, so tripling usage exhausts a balance rather than generating an invoice, and days you buy do not expire. AI tokens work the same way: a one-time plan bonus plus credit you choose to buy, with an optional workspace cost cap that warns or blocks at a threshold you set. Creating and updating tasks over MCP spends no AI tokens at all, so a busy agent is not automatically an expensive one.
Where does this fall short for a software team?▾
There is no git, CI, pull-request or code-review integration, so Onplana holds the plan and the work while your agent does the code work in its own environment and reports back. The phone companions are built but not generally available, iPhone is in TestFlight and Android on Play’s internal track. And hosted execution currently delegates to Claude’s managed agent sandbox, so a team standardised on a different provider runs agents from their own machine or the relay instead.
Which plan does a software team need?▾
Free covers the whole agent loop: MCP with two concurrent connections, the Issue Log, Ideas, comments, the review inbox and the audit trail, bounded by a one-time AI token bonus. Pro adds sprints, the Gantt with critical path, wikis and whiteboards, and includes five hosted agent-days a month. Business and Enterprise raise that to fifteen and forty, and hosted agent-days can be bought on any plan, including Free.

Point an agent at it and see

The free plan runs the whole loop: two agent connections, exclusive leases, the review inbox, and the audit trail behind all of it.

Start free See pricing

Already connected? Open Integrations to add another agent.