Everything it does, and everything it does not.

Install

Agentic PM is a desktop app. There is no terminal step and no developer tooling to buy.

The fourteen day trial needs no account and no card. Sign in only when you choose a plan at the end.

Your first run

Point it at the meeting notes you already have. Nothing else is required.

1. Open the app and create a workspace. Name it after the product or team, not yourself.
2. Drag in a folder of exported meeting notes, or connect a note taker below.
3. Wait for the ingest to finish. Fifteen meetings takes under a minute.
4. Open Slipped. That is the screen that matters.

Connect a note taker

Agentic PM does not record or transcribe anything. It reads what your existing tool produces.

How it works: five stages, and every one reports its own health

Most AI tools are a blank chat box. You paste a task, you carry the answer back by hand, and nothing about your product is remembered. Agentic PM is a pipeline instead: work enters on the left, passes through four stages, and comes out as an action that waits for you.

The part worth understanding is the last column of every node: a live health state. This is not a diagram of how it was designed. It is a map of what is running right now, and what is broken.

Senses: what comes in without being asked

Chat mentions, mail, calendar, meeting recordings, tracker state. None of this is something you fetch. It arrives, and the arrival is what starts everything downstream.

Reflexes: the parts that need no model

Sweeps, extraction, watchdogs, pre-meeting cards. These run on a clock and are deliberately mechanical, because the boring, structured majority of the week should not cost a model call or wait for your attention.

Brain: where judgment actually happens

The morning brief, the evening recap, meeting minutes, the weekly report. Model sessions are reserved for work that genuinely needs reasoning, which is what keeps this affordable on your own subscription rather than requiring a seat you buy from us.

Memory: what it knows between runs

Commitments, what you are waiting on, decisions, people, portfolio, and the context file that grows every time you correct it. This is the layer that makes the app know your team after six weeks without a setup interview, and the layer nobody else gives you.

Hands: nothing moves without you

Posting, documents, calendar, tickets, direct messages. Every node here is permanently gated. Not "gated by default" as a setting you could switch off. The approval is structural, which is why the state is its own colour on the map.

Where the commitment ledger sits

It is a reflex, not a model session, which is why it can run three times a day for free. It reads from senses, writes into memory, and when it finds something that slipped it hands the result to the brain to explain and to the hands to chase, with your approval.

The ledger

Every other tool stores the latest status of an action item. Agentic PM stores its whole history. That single difference is what makes a quietly broken promise visible.

// one commitment, as stored
{
  id:            "cmt_8f21",
  canonical:     "Ship the launch page",
  kind:          "commitment",
  owner:         "ops assistant",
  counterparty:  "founder",
  touches:       ["date", "dependency"],
  due:           "2026-06-25",
  status:        "open",
  history: [
    { meeting: "wk-2026-06-22", owner: "content lead", due: "2026-06-25" },
    { meeting: "wk-2026-06-30", owner: "ops assistant", due: null }
  ],
  flags: ["owner-changed", "silent-slip"]
}

An owner recorded as null is treated as a defect, not an empty field. It shows up as unowned rather than sitting quietly in a table.

The context file

One plain markdown file per workspace describing your product, your people, your vocabulary and the decisions already made. Every agent and every skill reads it before it starts.

You never sit through a setup interview. The file is written and extended as a side effect of ingesting meetings, and you can edit it directly at any time. After roughly six weeks it knows your team better than a new hire does.

The five things it watches for

Thresholds are configurable. load-risk defaults to 50% of open commitments held by one person.

Scope, date, dependency, money

Every commitment is tagged with what it touches. Anything touching none of the four sinks to the bottom of the list.

This is the line between a PM tool and a generic to-do app. A to-do app does not care that a date moved without a decision. A product manager is judged on precisely that.

Agents

Agents run on their own and watch things over time. Every one of these is in the build you can install today, with one exception that is marked as such.

The commands that ship with it

Skills are the things you invoke by name. Twenty three arrive in every new workspace, and any of them can be put on a schedule from the Routines tab without writing a config file.

You can add your own. A command is a markdown file in the workspace, so writing one is writing a document, and a command you delete stays deleted since 0.13.9.

Skills

Twenty three on-demand skills across six groups, included on every paid plan. Skills read the same context file the agents do, so output is about your product rather than a template.

Custom skills are plain markdown files you drop into the workspace. Yours are never uploaded.

CLI companion

The app is the product. The CLI exists for people who want to script it, and it is optional.

$ apm ingest ~/meetings/*.md
  read 15 meetings · 128 commitments · 41 open

$ apm check
  owner-changed    launch page      content lead → ops assistant
  silent-drop   balance transfer  last seen 22 Jun · touches money
  load-risk     10 of 16 open items on one owner

$ apm owed --to-me
$ apm standup
$ apm history cmt_8f21

Export and portability

Everything exports to plain markdown and JSON, on every tier including free. There is no export paywall and no proprietary format.

If you cancel, the ledger stays on your machine and stays readable. Cancelling removes the agents, not your history.

Privacy and data

• The ledger is local. It lives on your device, not on a server.
• Transcripts are never uploaded. Not for indexing, not for training, not for support.
• Commitment text is sent to a model. Matching the same promise across meetings needs language understanding, because the wording changes weekly. What is sent is the commitment line, not the meeting.
• Nothing trains anything. Your content is not used to improve any model.
• Team plans share context, not transcripts. The shared layer is the context file and the commitment ledger, both of which you can read and edit.

Known limits

1. It needs an API key or a paid plan for cross-meeting matching. There is no free lunch on the language model.
2. No time-saving figure is claimed anywhere, because none has been measured. When one exists, taken on real work with a readable method, it will be published with the method.
3. The evidence so far is one operator's own meetings. Fifteen meetings, one team. Enough to prove the failure is real, not enough to say how often it hits you.
4. The watchdog does not chase anyone yet. v1 nags inside the app only.
5. It is not a note taker and never will be. No recording, no bots in your calls.

FAQ

Changelog