Do you store our AI conversations?

No. The AI session context lives in a branch on your own repository. You own your IP, full stop.

Do we need to abandon GitHub?

No. Hervé sits on top of GitHub or GitLab — your repos, PRs, branch protection, and merge flow stay exactly where they are. You get a richer review view of the same PRs.

Are you an AI review agent?

No — and you can keep yours. Review agents like CodeRabbit, Codacy or Aikido check the code. Hervé shows you the thinking that produced it: the prompts, the plan, the rejected alternatives. Your coding agent already plans? Sure — but that plan dies in the session. Hervé ships it with the PR so the reviewer actually sees it.

Do you integrate with our stack?

Hervé works with GitHub and GitLab, and captures sessions from Claude Code, Cursor, GitHub Copilot CLI, OpenAI Codex, and OpenCode. Setup is one install script and one command per repo.

Where your team reviews AI-generated code

Agents write faster than you can review

Hervé is the review surface designed for the agentic era. The prompts, the code, and the process that generated it, in one unified page.

Start your free trial

14 days free · no credit card · one minute to set up

Devs who take code review seriously use Hervé

250

pull requests reviewed with Hervé

322h

of review time saved

The review surface

What you see when you open a PR

01 · change map

See what changed before you read a line.

Every file in the pull request, sized by how much it changed and coloured by the kind of risk it carries.

02 · criticality

Fourteen files changed. Only three are risky.

Start with those three. The rest you can skim, or skip — your review time goes where it counts.

03 · navigate in place

An IDE feel, without leaving the PR.

Click a symbol to see where it is defined and where it is used, across the files this PR changes. Press / to search symbols. No checking out the branch, no jumping between GitHub, your editor and Claude Code to answer one question.

04 · the session behind the PR

The prompts, the skills, the dead ends.

Set up Hervé to capture your AI sessions, and every PR arrives with the one that wrote it: what was asked, what the agent tried and dropped, every skill it used.

05 · anti-drift

Nobody asked for this.

When a change contradicts your instructions, or nothing you asked for explains why it is there, Hervé says so — and names the ticket and the prompts it checked against.

06 · ask on the line

Ask why a line was changed.

Tag @herve-review in the thread you are already in. It reads the linked issue and the session that wrote the code, and quotes what it found.

07 · across the repo

See how your team is prompting.

Hervé captures your sessions on commit, so you can see which skills get used, which files consume the tokens, and where the team is putting its prompting effort.

app.herve.review/github/gondor/beacons/pull/482

The session behind this pull request41 turns · 88k tokens

linked issue · #311

Beacons drop on 5xx. Add retry with backoff. Don't touch the dispatcher, it's shared with rohan.

ticket loaded

#311 · Beacons drop on 5xx

skill

/grill-me

plan · superseded

Move the backoff into the dispatcher so rohan gets it too

final plan

Retry wrapper around send(), exponential, capped at 30s

skill

/implement

implementation

edited 6 files, +451 −119

execution

pytest · 3 failed, then passed

skill

/code-review

3 skills invoked · 41 turns · 88k tokens

Change map14 files · 2,047 lines

Files

changed

Lines

2,047

added and removed

Critical + high

where your review hour goes

14 files · 2,047 lines · click any file to open it

Codesrc/beacons/sender.py

83@retry_on(HTTPError)

84def send(payload, *, retries=3):

85 for attempt in range(retries):

86 _backoff(attempt)

87 resp = _post(payload)

88 if resp.ok: return resp

89 raise RetriesExhausted(payload.id)

src/util/timing.py:12

12def _backoff(attempt: int) -> None:

13 """Exponential, capped at 30 seconds."""

14 time.sleep(min(2 ** attempt, 30))

click any symbol for its definition or callers · / to search symbols

Codesrc/queue/dispatch.py

linked issue · #311

Beacons drop on 5xx. Add retry with backoff. Don't touch the dispatcher, it's shared with rohan.

75class Dispatcher:

76 """Shared with rohan. Do not change lightly."""

77 def enqueue(self, job):

77 def enqueue(self, job, *, priority=0):

78 self._q.put((priority, job))

Hherve-reviewboton dispatch.py:77

This contradicts the planThe request says don't touch the dispatcher, it's shared with rohan. This changes its signature.

1 flag on this pull request · posted on the line it concerns

Conversationsrc/queue/dispatch.py

77 def enqueue(self, job):

77 def enqueue(self, job, *, priority=0):

PTPippinTookon dispatch.py:77

@herve-review why does this take a priority argument now? Nothing in the issue asks for priorities.

Hherve-reviewbotreplied in 12s

It came from a plan the session dropped. At turn 22 it proposed move the backoff into the dispatcher so rohan gets it too, added the argument, then went back to wrapping send() and never removed it. Issue #311 asks only for retry with backoff and never mentions priority.

answered in 12s · from issue #311 and the session that wrote the code

Insightsgondor/beacons · last 30 days

Commands & skills

invocations, and distinct checkpoints that used it

/tdd34 ×12

/code-review21 ×9

/pr-prep18 ×8

/implement9 ×4

/diagnosing-bugs2 ×2

Where the tokens go

per file, not per API key

beacons/sender.py412k

client/generated/api.ts288k

queue/dispatch.py143k

tests/test_retry.py96k

api/beacons.py31k

gondor/beacons · updated as your team's sessions are captured

Know what this PR touches in one second.

One surface to rule them all

Everything you just scrolled, on one screen — everything you need to judge an AI-generated PR.
You still read the code, you’re still in the driving seat.

change map · criticality

Fourteen files changed. Only three are risky.

The change map sizes every file by how much it changed and colours it by the kind of risk it carries.

_backoff(attempt)

src/util/timing.py:12

def _backoff(attempt: int):

sleep(min(2 ** attempt, 30))

navigate in place

An IDE feel, without leaving the PR.

Click a symbol and the definition opens where you are. Press / to search the symbols this PR touches.

PROMPT

“add retry with backoff”

SKILL /implement

PLAN · SUPERSEDED

Backoff inside the dispatcher

the session behind the PR

The prompts, the skills, the dead ends.

What was asked, what was considered and dropped, which skills were used — on every PR whose session your team captured.

− def enqueue(self, job):

+ def enqueue(self, job, *, priority=0):

This contradicts the plan

The request says “don’t touch the dispatcher”.

anti-drift

Nobody asked for this.

Hervé flags what you never asked for, and what contradicts what you did — naming the ticket and the prompts it checked against.

@herve-review why a priority arg?

From a plan the session dropped at turn 22. Issue #311 never mentions priority.

ask on the line

Ask why a line was changed.

Tag Hervé in the thread you are already in. Every claim in the answer quotes its source.

/tdd

/code-review

sender.py 412k

across the repo

See how your team is prompting.

Which skills actually get used, and which files consume the tokens, repo by repo.

Integrates with your stack

Get started

Up in one minute

Once installed, work as you normally would. Hervé hooks into git, captures the agent session, and ships it with your next PR. Your code and prompts stay safe on your repo.

Where your code and prompts live →

macOS / Linux

$ curl -fsSL https://app.herve.review/install.sh | bash

my_repo $ rv install

Windows

$ scoop bucket add herve https://github.com/ffelten/scoop-bucket.git

$ scoop install herve/rv

my_repo> rv.exe install

Requires Git for Windows and Scoop.

FAQ

Common questions

No. Your code stays on your git, hosted on GitHub or GitLab. Hervé reads from there and never copies or stores it on our servers. A breach of our database would leak zero lines of your code.

Still have questions? Drop us a line at hello@herve.review.

See the thinking behind your next PR.

Start your free trial

14 days free · no credit card · one minute to set up

Agents write faster than you can review

See what changed before you read a line.

Fourteen files changed. Only three are risky.

An IDE feel, without leaving the PR.

The prompts, the skills, the dead ends.

Nobody asked for this.

Ask why a line was changed.

See how your team is prompting.

One surface to rule them all

Integrates with your stack

Up in one minute

Common questions

Do you store our code?

Do you store our AI conversations?

Do we need to abandon GitHub?

Are you an AI review agent?

Do you integrate with our stack?

See the thinking behind your next PR.