Skip to content
Reality Graph

Comparison

Greptile vs. Local Verification

Last updated: 2026-11-038 min read

Sources checked 2026-11-038 sourced figures

A codebase index is a tool-maintained representation of your whole repository, built from embeddings or a graph, that lets a reviewer see relationships beyond the changed lines. Greptile reviews pull requests from such an index; local verification checks a change against its written task. The index answers questions about structure: what calls this function, which modules a change touches. It cannot answer whether the change did what the task said, because the task is not in the repository. That second question is what local verification exists for, and the index also widens the data question from the diff to everything.

“tests pass”reported by the agentexit code 0run by Reality Graphevidenceonly execution counts
Contents

What a codebase index answers, and what it cannot

The definition above is the site’s own, from the glossary. It names two properties at once. The first is reach: a diff shows the lines that changed, an index shows what those lines are connected to, so a reviewer backed by an index can notice that a renamed field is read in a module the pull request never opened. The second is scope: to offer that reach, the tool has to hold a representation of the whole repository, not only the change.

Structure is the class of question an index is good at. Which functions call the changed one. Which tests cover the changed path. Whether a similar function elsewhere validates its input and this one does not. These questions have their answer inside the repository, and a graph of the repository can find it.

The question an index cannot answer is whether the change did what the task said. A pull request that renames a field cleanly across all call sites can still be the wrong change if the task asked for a migration, not a rename. The reference for that judgment, the written task with its scope and acceptance criteria, is not in the repository, so no representation of the repository contains it. Reading code against code finds structural defects; reading the change against the task is a different check with a different reference, the distinction set out in code review versus verification.

That is the frame for this comparison: two questions, two references, two outputs. A codebase-indexed reviewer and a local task check are not two answers to one question but one answer each to two questions. The sections below place Greptile in that frame with its own documented facts, then follow what the index does to the data question, which is the part a procurement review tends to underweight.

Where Greptile sits in that frame

Greptile’s documentation describes an AI code review agent that parses every file of the connected repository into functions, classes and variables, maps their calls, imports and dependencies, and queries that graph on each pull request, posting findings as comments with suggested fixes (Greptile docs, checked 3 November 2026). Its pricing page (checked 3 November 2026) lists Starter free for one developer with 50 credits a month, Pro at 1 US dollar per credit without seat fees, and Enterprise at custom pricing, where self-hosting sits; a Base review costs 1 credit, a Plus review 3, an Apex review 10. The seat price of 30 US dollars this page cited in July 2026 is no longer offered.

What Greptile documents about where the code goes

The security page (dated 18 October 2026, checked 3 November 2026) states that Greptile Cloud runs on AWS, stores customer code on an encrypted filesystem until repository access is revoked in the code host, sends inference to third-party LLM providers under zero-data-retention agreements, may use de-identified data for AI training unless an organisation admin opts out, and deletes production data within 24 hours of an admin request, backups within 30 days. For self-hosting, the deployment documentation (checked 3 November 2026) describes Docker Compose or Helm on Kubernetes with your own model provider and lists outbound HTTPS to the model and code-host providers as a requirement; the security page adds that a self-hosted installation collects no data for machine learning unless you configure it.

Why the index makes the data question bigger

For a diff reviewer the data at stake is the change. For an indexed reviewer it is the whole repository, held in a representation that lives as long as the index does. That shifts the data question from which lines leave the environment to three others: where the representation lives, how long it is kept, and who may learn from it.

The documented facts above map onto those three questions without settling them. Where: a vendor filesystem and vendor-side model providers in the cloud, your own hosts in the self-hosted edition. How long: as long as repository access stands, plus a backup window. Who learns: a training setting that an admin has to turn off, and a different default when you host it yourself. Each line is a vendor statement about a default; the contract, the region, the configured model endpoints and the retention and access settings of your installation decide what is true for you.

Self-hosting narrows the first question and does not close the other two by itself. Outbound HTTPS to a model provider is listed as a requirement, so an air-gapped installation means a model endpoint inside the perimeter, and the index still exists as a second copy of the repository on hosts you now operate. The general method for drawing that line, which data may cross which boundary, is in the local AI code review guide; the comparison table below applies it to three setups.

Four questions settle the data side of any indexed reviewer, cloud or self-hosted, before a trial starts:

  • Where the representation lives: which filesystem, which region, and which model endpoints receive which part of it.
  • How long it lives: what revoking access deletes, what a deletion request deletes, and how long backups outlive both.
  • Who may learn from it: which setting governs training on your data, what its default is, and who in your organisation can change it.
  • What the installation talks to: every outbound connection a self-hosted instance needs, and which of them an air-gapped setup would have to replace.

Write the answers next to the vendor statements they test. A procurement review that holds both columns can say which statements it verified and which it accepted; one that holds only the vendor’s column cannot tell the difference.

The comparison, without the spin

Three setups, three reviewable boundaries

Cloud, self-hosting and a local task check answer overlapping but non-identical questions.
Greptile CloudGreptile self-hostedLocal task check
Review referencePull request plus codebase indexEditorial decision mapping · 2026-11-03The site's mapping; the vendor pages are the source of the facts.Pull request plus codebase indexEditorial decision mapping · 2026-11-03The site's mapping; the vendor pages are the source of the facts.Written task, scope, criteria, local checksEditorial decision mapping · 2026-11-03The site's mapping; the vendor pages are the source of the facts.
Documented data pathCode on the vendor's encrypted filesystem on AWS until code-host access is revoked; third-party model providersGreptile security statement · 2026-11-03Vendor statement; check contract and configuration.Code on your hosts; your model provider; outbound HTTPS to model and code hostGreptile deployment documentation · 2026-11-03Vendor statement; dependencies depend on configuration.Only the locally selected commands and artifacts; outbound services assessed separatelyEditorial decision mapping · 2026-11-03Product-neutral mapping; no completeness promise.
OutputPull-request comments with suggested fixesEditorial decision mapping · 2026-11-03The site's mapping; the vendor pages are the source of the facts.Pull-request comments with suggested fixesEditorial decision mapping · 2026-11-03The site's mapping; the vendor pages are the source of the facts.Observed check outcomes, bounded task decisionEditorial decision mapping · 2026-11-03The site's mapping; the vendor pages are the source of the facts.
Residual decisionDo cloud data flow, plan and review quality fit?Editorial decision mapping · 2026-11-03The site's mapping; the vendor pages are the source of the facts.Do operations, model provider, network and licence fit?Editorial decision mapping · 2026-11-03The site's mapping; the vendor pages are the source of the facts.Are the task and checks sufficiently appropriate and complete?Editorial decision mapping · 2026-11-03The site's mapping; the vendor pages are the source of the facts.
Cloud, self-hosting and a local task check answer overlapping but non-identical questions.

When Greptile is the right choice, and when the local check wins

Choose an indexed reviewer when cross-file effects are a real source of defects in your repositories and the documented data path, cloud or self-hosted, fits your constraints. Decide on a sample: representative pull requests, a rubric written before the trial, and a count of useful findings, false positives and misses. Architecture, certification and aggregate marketing results tell you nothing about that count.

Add a local task check when the open question is conformance: did the AI build what we asked? Code review and task comparison use different references, and a clean structural review of the wrong change is still the wrong change. Whether either check is sufficient depends on the written task, the selected commands, the observed output and the human decision at the end.

The two are not alternatives in the strict sense. Teams that run both keep two records per change: what the structure around it looks like, and whether it did what was asked. Teams that run only one should know which question they have left unanswered.

Where Reality Graph fits

Reality Graph occupies the local task-check slot in the table. It compares a governed run with its written task and records observed results in an evidence report. Its local boundary covers the commands and artifacts selected for that run; any configured external service must still be assessed separately. It does not establish Greptile’s quality, security or compliance, and it does not replace PR review. What that slot covers, and where it is still thin, is set out in the feature list and the gaps it admits.

This comparison provides

  • Dated vendor statements
  • Three distinct data paths
  • Reviewable procurement questions
  • Explicit residual decisions

It does not provide

  • A product ranking
  • A compliance or security verdict
  • A guaranteed detection rate
  • A claim that one setup replaces another
Compare the concrete configuration with your policy and task; do not assess the product label alone.

FAQ

What are the differences between Greptile and local verification?
They read a change against different references. Greptile reads it against a codebase index, a graph of the repository, and reports what the change does to the structure around it. Local verification reads it against the written task and the checks you selected, and records whether the change did what was asked. One produces review comments on the pull request, the other an observed result per run. Neither answers the other's question.
Does Greptile store my code?
In the cloud edition the index is the stored code: the repository sits on the vendor's encrypted filesystem for as long as repository access stands, and inference runs through third-party model providers (security page, checked 3 November 2026). The useful question is not whether the code is stored but where, for how long and who may learn from it. The data-question section above walks through those three; the self-hosted edition changes the first answer and leaves the other two to your configuration.
Can Greptile run on-premises?
Yes, in the Enterprise plan, as Docker Compose or Kubernetes with your own model provider (deployment documentation, checked 3 November 2026). On-premises is not the same as offline: outbound HTTPS to the model provider is a documented requirement, so an air-gapped installation needs the model inside your perimeter, and the index remains a second copy of the repository on hosts you now operate. Check model, network, storage and operational controls rather than the label.
What does Greptile's credit pricing mean for AI-heavy teams?
You pay per review and by depth, not per person: Starter is free with 50 credits a month, Pro costs 1 US dollar per credit, and a Base, Plus or Apex review costs 1, 3 or 10 credits (pricing page, checked 3 November 2026). For a team that opens many small AI-assisted pull requests, the bill follows review count and tier, so model your own volume instead of a seat count and recheck before procurement; the model changed between July and November 2026.
Is full-codebase context worth it?
It is worth measuring, not assuming. An index can surface relationships outside the diff, and whether that turns into useful findings in your repositories is an empirical question. Run representative pull requests and record useful findings, false positives and misses before you decide. Architecture establishes no detection rate; only your own sample does.
Can we combine Greptile with a local verification layer?
Yes, as two separate checks with two separate references. The PR reviewer reads the structure around the change; the local task check records how the change compares with the written task and the commands you ran. Keep the outbound services of each visible, and do not let a green result from one stand in for the other: neither check establishes that the other is complete or unnecessary.

Keep reading

Sources

Want to see what your last agent run would have looked like?

Request access