Build an always-on code reviewer

Code review is the thing small teams drop first. Not because nobody thinks it matters, but because it needs someone with context to be free at the same moment a pull request is open, and on a team of two or three that moment frequently does not arrive.

An agent with a clone of your repo, running on its own machine, that reviews when you message it, fills that gap reasonably well — provided you are precise about what you are asking it to do.

What an agent reviewer is genuinely good at

• The boring pass. Error paths that swallow exceptions, a new code path with no test, a config value read but never validated, a function that got long enough to be doing two things.
• Consistency with the rest of the codebase. It has the whole repo on disk, so "we do not do it that way here" is a check it can actually make.
• Being available. At 11pm on a Sunday, which is when a lot of small-team code gets written.

What it is not good at

Deciding whether the change was the right thing to build. Understanding the customer conversation that led to it. Knowing that this module is being deleted next month so the duplication is fine. Those need the context a human has and the agent does not, and an agent that confidently reviews them produces confident nonsense.

Treat it as a first pass that catches the mechanical problems so the human review can spend its attention on judgment. That framing is the difference between this being useful and being noise.

Which runtime

The choice comes down to what you want the output to be.

Aider is the natural pick if you want changes as well as comments. Its unit of work is a git commit, so a review that finds a missing null check can produce the diff that adds it — a diff you read and revert like any other, rather than a suggestion you have to implement yourself. $9 a month for the box.

Claude Code is the pick if you want depth of analysis over commits. It is the strongest tool user of the 9 and will explore the codebase further before forming an opinion. $9 a month.

deepagents suits a large or unfamiliar codebase where the review needs a plan before it starts — at a higher cost per review, because planning means more model calls.

Full comparison of all 9 is on the hosted coding agents page.

Setting it up

1. Describe the reviewer, including its limits. Say what to check and, just as importantly, what to leave alone. A reviewer briefed to "review the code" will comment on your naming conventions forever.
2. Give it repo access. A deploy key or a token scoped to the repo. Scope it as narrowly as the work allows — this is a credential on a machine that runs a shell.
3. Clone once. The workspace is on a persistent volume, so the clone survives restarts and idle stops. Fetch per review; do not re-clone.
4. Connect Telegram or Slack. Message it a branch or PR number, get the review back in the same thread.
5. Tune the brief for two weeks. The first reviews will be too pedantic or too shallow. The brief is the dial.

The brief is the whole product

A reviewer brief that works tends to look like this:

Review the diff on this branch against main. Flag: swallowed errors, new code paths without tests, unvalidated external input, and anything inconsistent with how the rest of this repo does the same thing. Do not comment on formatting, naming preferences, or architecture decisions that are already made. If you are unsure whether something is a problem, say so rather than asserting it. Finish with the three things most worth a human look.

The last sentence does more work than the rest. A ranked short list gets read; twenty equally-weighted comments get skimmed and then ignored.

Cost and the failure mode

Hosting is $9 a month, fixed, first agent month free. Model tokens are at provider cost, and review cost scales with diff size — a normal PR is cheap, a thousand-file refactor is not.

The expensive failure is a reviewer that gets stuck: it cannot find something, keeps searching, and burns tokens producing nothing. Give it an explicit give-up instruction, and watch cost per conversation on the Billing tab for the first few weeks. A retry loop on a coding agent is the most expensive thing that happens on this platform, and it is silent unless something is watching.