Forge

The trait that decides it is that your shell stays yours. Your aliases, your completions, your prompt and every plugin you have accumulated keep working, and the agent lives alongside them rather than replacing the session with its own interface. After a week you stop thinking about it as a separate program, which is the highest compliment a terminal tool gets.

Pick Aider if you prefer a clear boundary between your shell and the thing editing your code, and if the idea of an agent inside your interactive prompt makes you uneasy, trust that instinct.

The integration is the risk. A configuration step wires this into the shell that runs every command on the machine, and the same program executes commands against your working tree with no container between it and the filesystem. Two capabilities that would each deserve scrutiny separately are combined in the process you use most.

Read what the setup step changes before running it, and keep it off machines with production credentials. What it does right: it performs git operations, so its work arrives as commits that can be inspected and reverted rather than as anonymous edits.

The interesting property is the bound rather than the roles. Separating execution, planning and research lets each sub-agent carry only the material its task requires, so a long research detour does not enlarge the context every subsequent edit is charged against. Cost per useful edit stays closer to flat as a session lengthens, which is the failure mode of single-context agents.

What the record does not describe is how the plan is validated before execution begins, or how results return to the planner. No benchmark is published.

An open command-line tool with a services tier above it is the standard open-core wedge, and it works when the paid layer solves something individuals cannot: team state, shared context, spend control. None of that is described here, and nothing carries a number.

That absence usually means the company is still learning what customers will pay for, which is a fine place to be and a bad thing to depend on. Moat: none yet beyond the shell integration habit. Likely path: a team tier, or the tool becoming a hiring advertisement for the parent company. Position: watch, do not standardise.

The install is free and the spend lands on whatever provider account we already have, which is the arrangement I prefer because it keeps the meter somewhere I can already see it. The constraint is uniformity: the integration targets one shell, and I have sixty engineers who do not agree about anything, including that.

There is no unattended mode, so nothing here reaches our pipelines or produces a record anyone can audit. Support is a small company with no agreement on offer. Approved with conditions: individual opt-in, our existing provider keys, and no production access from a machine running it.

A single Rust binary from an install script, and I checked the script before piping it anywhere. The licence is Apache-2.0, so the source is mine to read and change, and the model layer takes my own key through OpenRouter, which means a few hundred options and no vendor deciding which one I get today.

It is an MCP client, so my servers are available without a shim. What I miss is a local endpoint in the documented model list, which for something this hackable is an odd gap. Everything else about it is built the way I would have built it.