Codebot

Automatic

Every pull request, reviewed

As soon as a pull request opens or gets new commits, Codebot reviews it: findings on the lines they are about, a summary that stays up to date, and a check with the verdict.

feat(billing): invoice endpoint #482
Codebot / Code reviewChanges requested · 1 blocking issue
src/billing/invoices.ts
27export async function getInvoice(c: Context) {
28+ const id = c.req.param("id");
29+ const row = await db.query.invoice.findFirst({
30+ where: eq(invoice.id, id),
trycodebotHighInvoice lookup isn’t scoped to the organization

Any signed-in user who knows an invoice id can read another organization’s invoice. Add eq(invoice.orgId, c.get("orgId")) to the where clause.

Reply @trycodebot fix this and I’ll push the fix.

31+ });
32+ return c.json(row);
33}

Findings ranked by severity

Every finding is critical, high, medium or low. You pick the lowest severity worth an inline comment; the rest stays in the summary.

Blocks bad merges, if you want

Fail the check while critical or high issues are open, then require it with branch protection or a branch policy.

Re-reviews that remember

On each push it focuses on what changed since its last review, resolves its own threads once they are fixed, and doesn't repeat itself.

Skips what you tell it to

Drafts, bots, chosen authors, base branches, title keywords, ignored files or oversized changes. A skipped check says which rule matched; comment @trycodebot review to get one anyway.

Built for large pull requests

Source files go first, lockfiles and generated code last, riskiest changes first. When something doesn't fit, or time runs short, the review says what it didn't read.

GitHub and Azure DevOps alike

Inline findings, summary comments, votes and statuses work the same on both hosts. On Azure DevOps it works without project admin rights, from a service account you choose.

On request

Mention @trycodebot anywhere to hand it work

Ask it about a pull request, have it fix a finding, or give it an issue to implement. Mention it like you would a coworker; it answers where you asked, and code changes arrive as commits or pull requests.

Pull requestacme/web #482
  1. M
    Maya

    @trycodebot fix this, and add a test for it

  2. trycodebot

    Pushed 3f9c2e1 to the branch: the invoice lookup is scoped to the organization, with a test. Resolved 1 thread.

LinearOPE-41 Export invoices as CSV
  1. S
    Sam

    @trycodebot can you take this? Same columns as the PDF.

  2. TryCodeBot

    Opened pull request #491 in acme/web: a CSV export next to the PDF one, with the same columns.

Discord#billing
  1. P
    Priya

    @trycodebot why did last night's invoice job fail? fix it if you can

  2. TryCodeBot

    It timed out on the largest organizations after #470 removed batching. Opened #493 to batch them again.

Fixes findings on the pull request

It commits to the pull request's own branch, resolves the threads it fixed and never force-pushes. The push gets a fresh review.

Implements Linear issues

Assign an issue to Codebot and it works on a branch, opens the pull request and links it to the issue. Each team or project can point at its own repository.

Answers your team in Discord

Tag it in a channel you turned on. It keeps a status message up to date while it works, then replies with the answer or the pull request.

Reads your CI, and looks things up

Ask why a check failed: it reads the checks and their logs. It also reads its own past runs and reviews, and with an Exa key it searches the web.

Fast answers, no sandbox

With fast mode on, questions about pull requests, issues, deploys or errors are answered in seconds. When a request needs the repository, it hands over to a sandbox with what it found.

Picks up where it left off

Follow-ups continue in the same workspace and branch, with the conversation so far, instead of starting over.

On a schedule

Routines for the chores nobody gets to

Tell Codebot what to do and when: every Monday, check the dependencies; every morning, triage new Sentry errors. Each run starts fresh in its own sandbox, on one repository.

  • Hourly, daily, on weekdays, weekly, or any cron expression, in your timezone
  • It sees what the last run reported, and posts only when something is new
  • Results go where you choose: a pull request, Linear issues in one team, or a Discord channel
  • It pauses itself after three failed runs in a row, and says why

Integrations

It can see your deploys, errors and databases, not only your code

Connect Vercel, Sentry and Neon once. Codebot uses them when you mention it and in routines; reviews stay on the code.

  • Connected once by an admin
  • Read only by default
  • Off for the agent with one switch
  • Never reads your rows

How it works

Set it up once. Nothing new to learn.

Reviews need no settings to get going. The only thing to remember is the mention.

  1. 01

    Connect your git host

    Install the GitHub App, or connect Azure DevOps with an access token. Reviews turn on for the repositories you pick.

  2. 02

    Open a pull request

    Codebot reviews it in its own cloud sandbox and answers on the pull request itself. No CLI, nothing to run locally.

  3. 03

    Mention it for more

    Reply @trycodebot on a pull request, or mention it in Linear or Discord, to get an answer, a fix or a new pull request.

Your standards

It follows the rules your repository already has

Codebot reads AGENTS.md, CLAUDE.md and a review-only .codebot/review.md before every review. It reads them from the base branch, so a pull request can't rewrite the rules it is reviewed by.

  • Instructions of your own, on top of the repository's
  • Comment only, approve when clean, or request changes on blocking issues
  • A pass or fail check on the commit, for GitHub branch protection or an Azure DevOps branch policy
.codebot/review.md
# Review guidelines

## Always flag
- Queries that aren't scoped to the caller's organization
- Secrets or tokens written to logs
- Migrations that lock large tables

## Never flag
- Formatting; Prettier owns it
- Missing tests in `scripts/`

## Context
Money is stored in cents as integers.
A rounding change is at least High.

In your control

A teammate you can trust with your code

You decide what Codebot may do, which model it uses and where it can be reached. It shows its work.

A sandbox for every task

Every review and request runs in its own isolated VM, or a container on your own runner. Reviews are read-only, and pull requests from forks are skipped unless you allow them.

Your models, your keys

Pick the model and bring your own key, or any OpenAI-compatible endpoint. Keys are stored as organization secrets, or kept on your runner.

You set what it may do

One set of rules everywhere: the model, your instructions, whether it may push, open pull requests or update Linear issues, and where it can be mentioned.

Everything in one activity feed

Every review and every request, wherever it came from, with live progress and a full record of what Codebot did.

Know what it costs

Compute by repository and workflow, and tokens, cache hits and model cost on your own keys, per model.

Limits per channel

In Discord, choose per channel which repository it works on, who may ask, whether it may change code, and a monthly compute cap.

Bring a key from

Anthropic OpenAI Google OpenRouter Mistral DeepSeek Groq

or any OpenAI-compatible endpoint: vLLM, SGLang, LiteLLM, Azure OpenAI

Self-hosted runners · Early access

Run it inside your network, on your models

Keep reviews and agent work on infrastructure you control: your machines, your keys, and models you host yourself.

Settings → Runners → Add runner
docker run -d --name codebot-runner --restart unless-stopped \
  -v /var/run/docker.sock:/var/run/docker.sock \
  -e CODEBOT_API_URL=https://app.trycodebot.dev \
  -e CODEBOT_RUNNER_TOKEN=cbr1_… \
  -e CODEBOT_OPENAI_COMPATIBLE='{"name":"Corp GPU","baseUrl":"https://llm.corp.internal/v1"}' \
  ghcr.io/simonepriuli/codebot-runner:latest

Runs on your machines

A runner is one Docker container on a Linux machine you control. Each review and request gets its own container there, and the repository is cloned there.

Models only your network can reach

Point the runner at GPUs behind your VPN. It lists the models they serve and Codebot offers them in the model picker; it never learns the endpoint's address or key.

Code and transcripts stay inside

The runner reads your git host and posts reviews itself. Codebot gets the run's steps, without your code, diffs or what the agent read and wrote.

Choose per repository

Send some repositories to your runners and keep the rest on Codebot's cloud. Runs on your runners don't count as compute on your plan.

FAQ

Questions, answered

What can I ask it to do?

Anything you would ask a teammate about the code: explain a change, fix a finding, apply its review, find out why CI failed, implement an issue. It answers where you asked, and code changes arrive as commits on the pull request or as a new pull request.

Where can I mention it?

In pull request comments on GitHub and Azure DevOps, on Linear issues (or assign the issue to it), and in the Discord channels you turn on. Each place can be turned on or off on its own.

How does it work in Discord?

Add Codebot to your server and turn on the channels it may work in, each with its own repository. Tag @trycodebot in one of them, use /codebot, or pick Apps → Ask Codebot on any message. It replies with a status message it keeps up to date while it works, then answers and notifies you; reply to it to continue. Per channel, you choose whether it replies in the channel or a thread, whether it may push or open pull requests, which roles may ask, and a monthly compute limit.

Can it work on a schedule?

Yes, with routines. An admin writes the request, picks a repository and when it runs, and chooses where results go: a pull request, Linear issues in one team, or a Discord channel. Each run sees what the last one reported and stays quiet when there is nothing new. A reply to a routine's Discord post continues the conversation.

Can it use Vercel, Sentry or other tools?

Vercel, Sentry and Neon, today. An admin connects them once in Settings → Integrations, and Codebot uses them when it is mentioned and in routines, never during reviews. Vercel and Neon are read only; Sentry only reads unless you switch on changes.

What is fast mode?

With fast mode on, Codebot first tries a request without a sandbox, using the tools that don't need the repository: pull requests and checks, Linear, its past runs and your connected services. Those answers come back in seconds. When a request needs the code, it continues in a sandbox with what it already found. Issues assigned in Linear and runs on your own runners always use a sandbox.

Will it push to my branches?

Only if you allow it. On a pull request it commits to that pull request's own branch, never force-pushes, and won't push to pull requests from forks. For a Linear issue or a Discord request it works on a new branch and opens a pull request.

Does it run my code?

Reviews are read-only: Codebot reads the diff and the checked-out repository, and doesn't install dependencies, build or run tests. When you hand it work it may run commands and tests to check its changes. Either way it works in an isolated sandbox in Codebot's cloud, or in a container on your own runner.

Can it run inside our network?

Yes, with self-hosted runners (early access). A runner is a Docker container on a machine you control: runs execute there, call your models from there, and post to your git host from there. Your code, diffs and agent transcripts don't pass through Codebot. You choose per repository whether runs go to your runners or Codebot's cloud.

Can it block a merge?

If you want it to. Turn on “Fail the check on blocking issues” and require the Codebot check in branch protection (GitHub) or a branch policy (Azure DevOps).

Which models does it use?

The one you pick, with your own API key from Anthropic, OpenAI, Google, OpenRouter, Mistral, DeepSeek or Groq, or any OpenAI-compatible endpoint. Keys are stored as organization secrets; with a self-hosted runner they can stay on the runner, and models only your network can reach show up in the model picker.

How much does it cost?

It's free during the private beta. Each organization gets sandbox compute hours every month, and the usage page shows exactly where they go, along with tokens and model cost on your keys. Fast mode answers, which skip the sandbox, count at 1.5x their time. Runs on your own runners don't use compute hours. We'll share pricing well before the beta ends.

Add Codebot to your team

Codebot is in private beta. Tell us about your team and we'll send you an invite; reviews start with your next pull request.

Request access

Have an invite? Sign in