Pi Agent

The deciding trait is that trying it costs you nothing. It reads the data directory the command line tool already uses, so your past sessions and your existing logins are present the first time you open it, and going back means closing the window. That is the rarest property in tooling: a change you can undo.

You are the wrong buyer if you do not already use this particular agent, because everything here is built around it. Pick it if you do and you want a window. Pick the terminal if it never bothered you.

The packaging is the risk. A web server, a language runtime and a copy of the agent SDK are bundled inside the download, which makes installation easy and makes the version you run somebody else's decision. When the underlying agent ships a change, this application ships it later, or not, and the row records no update mechanism that would tell you which.

What it does right is show the working. Thinking, tool calls and compaction state are visible while a turn runs, rather than after it.

1. Displaying context consumption and cost during a run turns two hidden quantities into observable ones, which is what allows a user to form any model at all of why a session degraded.

2. Branching from an arbitrary earlier message converts the conversation from a linear transcript into a tree, so an unproductive path can be abandoned without discarding what preceded it.

3. That second property is the more consequential: it makes exploration cheap and makes a wrong turn recoverable, which is the closest thing to experimental control this category offers.

359 stars, one author, and a total addressable market defined as people who already chose a different project. That is a derivative position: growth requires the upstream to grow, and the upstream shipping its own interface removes the reason to be here at all.

Moat: none, and the switching cost is zero in both directions by design. Likely acquirer: none. Likely path: it thrives while the upstream neglects the interface and stops when it does not. Position: free and pleasant, and not a thing to plan around.

Financially trivial: nothing per seat, and the spend belongs to whatever provider accounts my engineers already have. There are installers for all three desktop platforms, which my endpoint team can push through the usual channel without a new process.

Operationally it is sixty separate installations with no central configuration, no SSO, no audit log and no retention policy, and nothing that runs unattended in a pipeline. Onboarding is minutes. Approved with conditions: managed installation, and no expectation that I can report on what it does.

Reading the existing data directory instead of inventing its own is the decision I respect. My sessions and logins stay in the place I already back up, and this application is a viewer rather than an owner of that state, which means uninstalling it costs me nothing.

There is no MCP in either direction, so the tools available are the ones the agent underneath brought, and the capability surface is not mine to extend from here. MIT keeps the fork open, and the code is small enough to patch when I want a column changed.