Traycer

The deciding trait is that it keeps the agent you already pay for. It plans, then hands the work to Claude Code or Codex or whichever you run, so nothing about your existing setup has to change and no second subscription appears. If your sessions fail because the model started coding before it understood the task, this is aimed precisely at that.

What it adds is a step and a surface, which is real friction on small tasks. Pick it if your work is large and underspecified. Pick your agent's own planning mode if it is neither.

An unexecuted plan is a confident document. Nothing here compiles, runs a test or touches a file, so a specification referring to a function that was renamed last month looks identical to one that is correct, and the agent receiving it will attempt both with equal enthusiasm. The failure surfaces downstream, in a tool that will be blamed for it.

What it does right is preserve context across the handoff. A shared filesystem, artifacts and history travel with a task, so the receiving agent is not reconstructing intent from a paragraph.

1. Most agent-to-agent architectures pass a message and lose everything that produced it, so the receiving agent works from a compression of the sender's reasoning. Carrying artefacts and history alongside the task makes the handoff lossy by choice rather than by construction. 2. Allowing an agent to ask a question back is the other half of that: an ambiguous instruction has a resolution path other than guessing.

2. No evaluation is published, and the claims here are structural rather than performance ones.

Giving away the tier that drives your own subscriptions is smart positioning: it costs the vendor nothing in inference, builds the install base, and reserves the paid ladder for people who want the credits handled for them. Two years of shipping and an install base in the tens of thousands is a real business, not a repository.

Moat: the coordination habit, which is stickier than a model wrapper because it holds your task history. Likely acquirer: an editor vendor or a model vendor wanting the planning layer. Position: the credible closed-source option in this class.

Two thousand four hundred a month is a number I can put in a budget, and the included credits are the number I cannot, because consumption climbs with task size and nothing here forecasts it. The free tier that uses our existing agent subscriptions is the version I would actually pilot.

Procurement gets little: no directory integration, no provisioning, no audit trail of what was planned against which repository, and a desktop application limited to one operating system. Nothing runs unattended. Approved with conditions: the free tier only, on the extension rather than the desktop app.

I want to dislike this more than I do. The source is closed, so the planning prompts are theirs and a fork is impossible, and there is no MCP client to bring my own servers into it. That is the usual list and it is disqualifying for most closed tools.

What saves it is where the execution goes. A local runtime is a supported target and so is an agent I wrote myself, which means the part that touches my code can be entirely mine while only the planning is rented. Grudgingly, that is a defensible division of labour.