Private betaInvite only

The reviewer that learns your team's bar.

Modak is an AI code reviewer for GitHub. 8 specialists review every PR in shadow, calibrate against your humans per path, and earn auto-merge one path at a time. You get a dashboard that shows exactly where its judgment tracks your team, and exactly where it doesn't.

30-sec GitHub App install · priced per repo, not per seat · zero-retention API, we don't train on your diffs
modak · acme-corp/api · calibrationupdated 4m ago
 
docs / routine
features
sensitive
forbidden
UX
0.94
0.88
0.71
ARCH
0.91
0.86
0.74
SEC
0.89
0.78
0.58
IMPL
0.93
0.87
0.76
TEST
0.90
0.85
0.72
OPS
0.88
0.79
0.54
docs/routine readme, tests, lintfeatures product codesensitive payments, auth-adjacentforbidden never auto-merge
3
beta teams
420
PRs reviewed in shadow
0
auto-merges. by design.
The problem

Senior reviewers carry the bar. That bar is specific — to your repo, your paths, your failure modes. Most review tools ignore that and comment like generalists.

The wedge

Modak learns per-persona × per-path-tier trust from your team's actual approvals. A call on an auth migration and a call on a README typo are scored separately, against separate humans.

The promise

Ship only the narrow claim we can back: a review-quality dashboard that earns auto-merge. Not a bold auto-merger pretending the math works on day one.

How it works

Shadow first. Dashboard second. Auto-merge when it's earned.

Three honest phases. You never leave one until the calibration on the previous one tells you it's safe.

Week 1–4 · start here

Shadow review, labeled by your team

Install the GitHub App. Modak reviews every PR silently — no comments, no noise. Each human approve/request-changes becomes a labeled data point against the persona verdicts. Your workflow doesn't change at all.

∼120 labeled decisions on a 30 PR/week repo
After 30 labels · per tier

Calibration dashboard you can sell internally

Per-persona × per-path-tier agreement rates. Where Modak tracks your team; where it doesn't. Which persona is most often wrong on which paths. The artifact you hand to your staff engineer before asking for a merge permission.

below 30 on a tier, Modak holds — no drift on noise
After trust is earned

Auto-merge on the paths you trust

Path-by-path opt-in. Flip auto-merge on docs/ first. Maybe features next quarter. FORBIDDEN paths never unlock regardless of confidence. Every auto-merge writes a revertible audit trail and feeds back into recalibration.

earned per path, never globally
Modak vs CodeRabbit, Greptile, Bugbot

AI code reviewers, compared on the thing that matters.

AI review is crowded. Generalist comment bots are table stakes. The novel thing — the thing nobody else has shipped honestly — is a per-team, per-path trust surface you can point at.

ModakGeneralist review botsBundled IDE reviewers
Review model8 specialists, voted per path tierUX, Arch, Sec, Impl, Test, Ops, Prod, CompSingle generalist commentone voice per PRInline hints at edit-timeoptimized for author, not reviewer
Trust calibrationPer-persona × per-path-tier, updated from human labelsdashboard is the productGlobal severity thresholdsuser tunes configNone exposedvendor tunes internally
Decision powerShadow → opt-in auto-merge per pathearned after calibration clearsComment onlysecond queue for humansSuggests at edit-timeno gate role
Pricing$49/mo/repo, flatscales with services, not headcount$30–200/seat/moscales with headcountBundled / free tier"good enough" for low stakes
Install surfaceGitHub App, 30 secno SDK, no IDE, no MCPGitHub App, minutesoften requires CI wiringIDE plugin + accountper-developer rollout
Calibration

How we actually earn the right to auto-merge.

Post-merge signalweight
HUMAN_LABELApprove / request-changes from a human reviewer in shadow mode. Primary bootstrap signal.1.00
REVERTGit revert within 14 days of merge. Strong negative — but rare, so never used alone.0.85
HOTFIX_TOUCHSubsequent PR touches same files within 14 days and lands under label hotfix / fix.0.55
CI_MAIN_FAILThe merge commit or its immediate successor fails CI on main.0.40
ERROR_SPIKEOptional Sentry / DataDog integration — error rate on merged paths spikes within 48 hours.0.35
Window: 28-day rolling, recalibrated weekly, not daily.  ·  Hold branch: if n_labels < 30 on a path tier, Modak holds the previous threshold. No drift on noise.

Instead we weight five post-merge signals over a rolling 28-day window, with human approvals as the primary bootstrap. When signal is too sparse we explicitly hold rather than recalibrate on noise.

Credit for every bad merge is assigned per persona, per path tier. If a revert fires on a sensitive/ change, Security's weight on sensitive/ drops; their weight on docs/ doesn't. That's the actual unit of learning.

Honest about the hard parts

We'll be worse than our competitors for your first 90 days.

Cold trust weights mean the calibration matrix is mostly grey. The dashboard will look under-populated until roughly the 30-label-per-tier mark. In that window, Modak is a shadow-mode review bot that's slightly worse than the incumbents — because it isn't pretending to be calibrated when it isn't.

That's the tradeoff, stated out loud. We'd rather ship a tool that earns the right to auto-merge than one that ships auto-merge on day one and hopes the math works. If you want a generalist commenter today, our competitors are fine. If you want a reviewer that fits your team's bar in 90 days, install us in shadow mode now.

Sample shadow-mode review

What the dashboard surfaces, before any auto-merge is on.

Nothing posts on the PR in shadow mode. This view lives on your Modak dashboard alongside the human reviewer's actual decision — the comparison is the point.

Modak · shadowPR #482 · feat: retry on checkout failurereviewed 4 minutes ago
shadowWould request changeshuman reviewer: request changes · path tier: features
8 personas weighed in. 6 approve, 1 concern, 1 wants changes.

Modak's verdict and the human's verdict matched on this PR — both said request changes. That's one more labeled agreement point on the features path for this repo.

If this were live auto-merge on features (it isn't)

Trust-weighted score 0.74 below the repo's learned threshold T = 0.85 for features → Modak would escalate, not merge. This is the behavior you'd want, and the agreement matrix above confirms it.

UX0.42
SEC0.71
ARCH0.88
IMPL0.89
TEST0.86
OPS0.94
PROD0.90
COMP0.87
UX specialistconf 0.42 · path tier: features
2 high — accessibility regressions on icon button & error state
Changes requested
Missing aria-label on icon-only button

The new icon button at line 42 has no accessible name. Screen readers will announce "button" with no context.

// components/CheckoutButton.tsx:42
<button onClick={handleRetry}
+  aria-label="Retry payment">
  <RefreshIcon />
</button>

The human reviewer flagged the same issue in their comment at 15:42 — this is a labeled agreement point.

Security specialistconf 0.71 · path tier: features
1 medium — unbounded retry could amplify a downstream outage
Concern raised

The new handleRetry re-invokes checkout() on every click with no attempt cap. Back-end rate-limiter would catch abuse, but worth a local cap at 3 attempts and a short cooldown before the next high-traffic window.

Human reviewer did not flag this — recorded as a disagreement. Security's weight on features drops a touch; the matrix cell updates on next weekly recalibration.

6 more personas approved: Arch, Impl, Test, Ops, Prod, Comp. View full decision ↗

Trust & safety

Guardrails that don't bend.

Auto-merge is earned per path, never globally. These rails sit above the earned-trust layer — they're non-negotiable regardless of how green the calibration matrix gets. Clean legal story, clean data story, clean revert path.

forbidden
Auth, migrations, secrets → never auto-mergeConfigurable path globs. Modak reviews, a human clicks merge, full stop.
zero-retention
Your diffs don't train anythingZero-retention terms with Anthropic on every model call. Your code is not used to improve our personas, our calibration, or anyone else's repo. That's contractual, not vibes.
pii
PII redaction before any model callEmails, tokens, customer identifiers scrubbed before the diff reaches any model. Redaction report visible in audit log per PR.
loop-guard
Escalate after 3 rounds without convergenceNo infinite "please fix this" loops. Hands back to the operator with a summary.
audit
Every decision → structured JSON logPR, personas, per-persona × per-path-tier weights, threshold, verdict, operator action. Stream to SIEM on Team+.
matrix
Open review-decision matrixPublished in docs — exactly when Modak merges, escalates, or blocks. No hidden override.
Pricing

Per repo, flat. Scales with services, not headcount.

Free tier is permanent on one repo in shadow. Pro unlocks additional repos, path-by-path auto-merge, weekly recalibration, and the per-persona × per-path-tier dashboard.

Free
$0forever · 1 repo

Shadow mode on a single repo. The right way to evaluate.

Start free
  • 1 repo, shadow mode only
  • All 8 personas
  • Calibration matrix (after 30 labels)
  • 7-day audit retention
Team
$199/mo · up to 10 repos

+$19/mo per additional repo. For orgs with many services.

Start Team
  • Everything in Pro
  • Custom path-tier rules
  • Per-repo persona opt-in/out
  • SAML SSO
  • Audit stream to SIEM
  • 365-day audit retention
  • Existing Collective clients: ask about bundled pricing.
Enterprise
Custom

SOC 2 Type II, air-gapped option, custom DPA.

Contact us
  • Air-gapped deploy option
  • Custom persona tuning
  • SOC 2 Type II report
  • 99.9% uptime SLA
FAQ

Frequently asked questions about Modak.

What is Modak?

Modak is an AI code reviewer for GitHub. It installs as a GitHub App, reviews every pull request in shadow mode using 8 specialist personas (UX, Architecture, Security, Implementation, Test, Ops, Product, Compliance), and calibrates its verdicts against your team's actual approvals on a per-persona × per-path-tier basis. Auto-merge is a per-path opt-in, earned after the calibration matrix clears a threshold.

Shadow mode is the default and the permanent free tier. Nothing is posted on your PRs until you explicitly opt in, path by path.

How much does Modak cost?

Free forever on one repo in shadow mode — all 8 personas, calibration matrix after 30 labels, 7-day audit retention. Pro is $49/repo/month with path-by-path auto-merge and weekly recalibration. Team is $199/month for up to 10 repos (+$19/repo beyond) with SAML SSO and SIEM audit stream. Enterprise is custom and includes SOC 2 Type II and an air-gapped deploy option.

Will Modak work on a small repo with 30 PRs per week?

Yes, because the primary signal is human approve/request-changes under shadow mode, not reverts. On a 30 PR/week repo you generate ~120 labels/month against persona verdicts. The matrix passes its hold-threshold (30 labels per path tier) in 4–8 weeks for the tiers that see regular traffic.

For very quiet path tiers — say, migrations/ on a repo that ships one migration a quarter — the cell stays grey until signal accumulates. We don't lie about convergence; the dashboard just says "insufficient signal, holding."

How is Modak different from CodeRabbit, Greptile, or Bugbot?

Generalist AI comments are table stakes and the market is already crowded. We're not the best-funded entrant there. The defensible surface is per-team, per-path trust — a calibration artifact nobody else ships, because auto-merge is a liability unless you've done the math.

If you only want comments on PRs, one of those tools is probably fine. If you want a reviewer whose judgment actually closes on your team's bar, that's us.

Does Modak train on my code? What's the legal story?

Model calls use the Anthropic API under zero-retention enterprise terms — nothing from your diffs is retained by the model provider and nothing is used for model training. Separately, we (ProductLove, Inc.) do not use your diffs to train or improve our persona prompts, our calibration model, or anything shared across customers. Your per-repo trust matrix is yours.

PII redaction runs on every diff before the model sees it. SOC 2 Type II is available on Enterprise; we can execute a DPA and a mutual NDA before install.

What if a revert happens after auto-merge?

The revert signal is attributed per-persona × per-path-tier. The personas whose confidence most contributed to the bad call get their weight on that path tier docked. The repo's threshold on that tier rises on the next weekly recalibration. Nothing else changes.

Independently: if revert-rate on any auto-merge-enabled path exceeds 2% in a rolling 28-day window, Modak flips that path back to shadow automatically and posts an incident to your Slack channel. The operator re-enables it manually, not the tool.

Do I have to turn auto-merge on to get value?

No. The calibration dashboard alone is sellable — "here's where our AI reviewer tracks our senior engineers, broken out by path" is a useful artifact for any engineering org, independent of auto-merge. Plenty of teams will live on Pro forever with every path in shadow.

How is Modak related to The Collective?

Modak is part of The Collective — a ProductLove, Inc. product. It's designed to be adopted standalone, without using the rest of the Collective platform. We prove this by using Modak at The Collective exactly the way you'd use it: we install the GitHub App, sign up on modak.sh, pay per repo, and get value from the calibration matrix.

If you're already a Collective client, bundled access is available on Team and Enterprise plans.

Can it review Dependabot / AI-authored PRs?

Yes. Bot-authored and AI-authored PRs go through the identical decision path. Modak doesn't know or care how a diff was written — it reviews the diff. Dependency bumps that don't touch FORBIDDEN paths and pass persona review will be auto-merge eligible on the same per-path terms as human PRs.

Install in 30 seconds

Start in shadow. Earn the merge. On your paths, your bar, your timeline.

The GitHub App is free forever on one repo. Nothing posts on PRs. Your calibration dashboard lights up as your team reviews. Flip auto-merge on — path by path — when you're convinced.