Codebuff

Codebuff's idea is a team of subagents from one prompt: pickers, searchers, editors and reviewers working a task in parallel, which is pleasant to watch and occasionally faster than one agent doing it in sequence. The trait that decides it is the bill: subscriptions start at $100 a month for base usage, with no bring-your-own-key and no local model, so the meter is theirs and the ceiling is theirs.

Pick it if orchestration itself is what you want to pay for and you have watched it beat a single agent on your code. Pick OpenCode or Claude Code otherwise; both do the job for less and show you the bill.

Start with the meter. $100 a month buys 1x usage, and the page does not say what 1x is, so the number you are budgeting against is undefined. No BYOK, no local models, so there is no route around it. Then the architecture: multiple subagents, each with terminal execution, and no Docker sandbox, which means several processes with shell access that you did not individually approve.

Run it in a container and ask support what 1x means before paying. What it does right: the repository is Apache-2.0 and readable, and the SDK is callable from CI, so the runs can be bounded by something.

The documented architecture is an orchestrator dispatching specialized subagents. 1. The reviewer subagent verifies by another model's reading, not by running tests, so verification is an opinion rather than an execution. 2. There is no sandbox around any of them. 3. No benchmark is published, so the claim that decomposition improves outcomes is asserted, and the counter-hypothesis, that a single agent with the same context does as well, is untested.

The observation: coordination has a token cost, since every handoff re-sends context, and that cost is not on the page. What would change the assessment is a comparison against one agent on one task set.

Codebuff is running two monetization experiments at once. The paid ladder, $100 for 1x, with no BYOK, is a company that must take margin on every token, which only works if the orchestration is worth the markup. Then Freebuff: the same Apache-2.0 agent, ad-supported, and the repository was renamed to freebuff, which tells you which experiment the founders think has legs. Ads in a terminal is a bet that attention is worth more than seats.

No obvious acquirer; the orchestration is open and the free variant undercuts the paid one. Position: try pay-as-you-go, do not build a workflow on it until one model wins.

The demo, subagents dividing a task, is good. Procurement: the entry subscription is $100 a month for 1x usage, so sixty seats is $6,000 a month, and the multipliers above it are unforecastable because 1x is undefined. There is no BYOK, so we cannot route through the model contract security already approved, which means a new data-processing review for a vendor with no published retention terms. The free path is ad-supported, a conversation I will not have with legal.

Onboarding is an npm install. Not yet: BYOK or a defined unit, and retention terms in writing.

The repo is Apache-2.0, and I can read the orchestrator and the SDK, which is more than most closed tools give me. Then the wall: no BYOK, no local models, just their credits at a cent each or $100 a month. The open code is a client for a closed service, the pattern I hate most, because a fork gives me an orchestrator that talks to nothing until I rewrite the model layer myself.

That rewrite is possible, and someone will do it, which is the only reason to score longevity above the floor. Grudging respect for opening the orchestrator; none for the rest.