Set your own coding standards, Built-in security scanning, Works with Claude Code and Cursor, Free to start, unlimited seats, Runs in your terminal, Use your own API key.
AI code reviewer that knows what your company does. It checks security in every review, shares rules with coding agents, has no per-seat fees, reviews locally before you push and has a self-hostable open core.
SourceAnt learns your team’s decisions, conventions and constraints from real pull requests, and applies them in every review. Each comment says why the change matters here.
Free for 3 repositories. Unlimited seats. Your own model keys.
Works with
- GitHub
- Claude Code
- Codex
- Cursor
250
free events every month
$0
markup on model tokens
Unlimited
seats on every plan
MIT
open source core you can self-host
Agents write the diff. Nobody wrote down why.
A change can pass every test and still undo something your team settled months ago. The reason lives in an old pull request, a review thread, a rule someone remembers.
SourceAnt keeps that reasoning, keeps it current, and brings it to the line that breaks it.
01 · Review
Comments that cite what your team decided.
Every pull request is read against its callers, your conventions and your organization’s rules. A finding names the risk, the rule it breaks and where that rule came from.
- Callers and contracts outside the diff, from the code graph
- Nitpicks off by default; you choose how much reaches the team
- Review across repositories grouped into one system
_quote escapes " but never \, so any review text carrying a backslash (a regex, a Windows path) is emitted as "...\d+", and a value ending in \ swallows the closing quote. The reader gets a malformed string that the Codetown parser can reject, losing the whole town. Escape backslashes before quotes.
Suggested change
def _quote(text: str) -> str:
escaped = " ".join(str(text).split()).replace("\\", "\\\\").replace('"', '\\"')
return '"' + escaped + '"' 02 · Security
Security checks in every review.
Semgrep reads the changed files in more than thirty languages before a model sees the change. Its findings reach the review as fact, and the model spends its pass on what a tool cannot catch.
- Semgrep: today
- OWASP ZAP: planned
- Nuclei: planned
- Trivy: planned
- Gitleaks: planned
03 · Knowledge
Learned from real work. Kept current.
SourceAnt proposes decisions and conventions from merged pull requests and pushes. You accept them once; reviews apply them from then on. When a rule is replaced or stops showing up, it is marked superseded or stale, never silently kept.
Memory keeps what happened. Knowledge keeps what your team holds true.
-
04 · Organization
Set a rule once. Every repository follows it.
Organization rules apply to every repository and workspace you own, not only where someone remembered them.
Charges never run in a request handlerRule
-
05 · Memory
The history behind a change.
Pull requests, reviews, pushes and every change to your knowledge, with when, why, who and how it went.
#391 merged, convention acceptedJul 02 -
06 · Local
Review before you push.
Ask your coding agent to review the working tree. Findings stay on your machine until you decide.
● review_working_tree
HIGH charge.ts:43 double charge
For agents
Your agents read the same rules your reviews apply.
Claude Code, Codex, Cursor and any MCP client query your team’s knowledge while they write, so fewer findings reach review at all.
codex · payments-api
› Add a refund endpoint.
● sourceant · knowledge for src/billing
Convention retries go through the queue
Convention amounts in minor units
Constraint no charges in request handlers
Writing refund.ts with a queued, idempotent refund.
Compare
Keep the reasons behind your code.
CodeRabbit and SourceAnt review pull requests across repositories, but they organize team knowledge and pricing differently.
CodeRabbit may fit if its reviews and learnings already meet your needs.
Compare the details| Area | CodeRabbit | SourceAnt |
|---|---|---|
| Knowledge | Stores review learnings and repository instructions. | Stores decisions, constraints and conventions linked to code and each other. |
| Agents | Learnings guide its own pull request reviews. | Claude Code, Codex and Cursor read the same decisions over MCP. |
| Repositories | Reviews linked repositories, with allowances depending on the plan. | Groups repositories into systems and reports review access gaps. |
| Pricing | Offers per user plans with features depending on the tier. | Unlimited seats, plans from $14 a month, plus your own model costs. |
Start
Your first review in three steps.
01
Install
Add the GitHub app to your organisation and pick the repositories.
02
Review
Open a pull request. SourceAnt comments on the lines that need attention.
03
Teach
Accept a decision once. Every later review applies it.
Proof
Reviews you can read in public.
Each is a public pull request in SourceAnt’s own repository, reviewed by SourceAnt.
- #210 feat(reviews): Change impact, Codetown towns and whole-file reads
- #211 feat(work): GitHub issue and pull request tracking
- #212 fix(codetown): Load bundled vocabulary in compiled builds
- #213 feat(mcp): Token scopes and revocation
- #214 fix(reviews): Run previews in cancellable jobs with bounded model work
Pricing
Pay your model provider directly. We add nothing on top.
One event is one webhook received. Redeliveries don’t count. MCP reads are free.
-
Free
$0
- 3 repositories
- 250 events a month
- Pauses at the limit
-
Solo
$14/mo
- 10 repositories
- 2,000 events a month
- $9 per extra 1,000
-
Starter
$49/mo
- 25 repositories
- 5,000 events a month
- $8 per extra 1,000
-
GrowthMost popular
$249/mo
- 150 repositories
- 40,000 events a month
- $6 per extra 1,000
-
Scale
$1,200/mo
- Unlimited repositories
- 200,000 events a month
- $3 per extra 1,000
Every plan: unlimited seats, your own model keys, pause or bill overage up to a cap you set.
Compare plansFAQ
Questions teams ask first.
Can I review changes before committing or pushing?
Yes. Open Reviews in the local app, or ask your coding assistant to call review_working_tree through local MCP. Requesting findings uses your configured model provider.
Can I use my own model keys?
Yes. Every review runs on your own model provider, which bills you directly. The open source core can also run on your own infrastructure.
Which security scanners run today?
Semgrep, inside every code review. OWASP ZAP, Nuclei, Trivy and Gitleaks are planned and not available yet.
How is Memory different from Knowledge?
Knowledge is what your team holds true: decisions, conventions and constraints. Memory is what happened: the events that led to it.
Which code hosts are supported?
GitHub.
Get your first review today.
Install the GitHub app, open a pull request and read what SourceAnt finds.
SA sourceant High
A gateway timeout after the charge was accepted retries it, so the card is charged twice.
Convention Payment retries go through the queue with an idempotency key. Learned from #391
Suggested change
return paymentQueue.enqueue(order, { key: order.id })