Self-hosted AI code review · Governance-first · BYO-LLM
Gatekeep is an AI code reviewer that runs entirely inside your own infrastructure. Every pull request gets reviewed by your own LLM, and merges are gated by a compliance-grade check with an immutable audit trail. Built for teams whose code can't go to a third-party SaaS.
Add Stripe refund endpoint #4127
feat/refunds into main · 6 commits · 3 files changed
Merge blocked
Override requires written justification — logged to the audit trail
AI now authors a large share of production code — and it ships measurably more defects than humans do.
Sources: issue and security-bug rates from CodeRabbit's published analysis of 470 open-source PRs (10.83 vs 6.45 issues/PR); adoption and approval-bias figures from industry surveys aggregated in our July 2026 market write-up.
Your PR queue is filling faster than any human team can read it. The obvious fix is an AI reviewer — but there's a catch that stops regulated and IP-sensitive teams cold:
CodeRabbit, Greptile, Ellipsis, Copilot Code Review — the effective tools route your code, or its metadata, out to multi-tenant OpenAI and Anthropic APIs. If you're under HIPAA, PCI-DSS, SOC 2, the EU AI Act, or a defense air-gap mandate, that's a non-starter. Your source is proprietary IP that cannot leave your perimeter.
Most tools auto-publish comments a developer can ignore. Even "blocking" checks are usually bypassable with a single checkbox — no justification, no record. Surveys find ~96% of developers don't fully trust AI-generated code, yet roughly half don't consistently verify it before committing (Sonar, State of Code 2025). That's not a merge gate. That's noise with a green checkmark.
So regulated teams are stuck choosing between shipping unreviewed AI code and not using AI review at all.
There's a third option.
The white-space no incumbent covers: real self-host, your own model, an enforceable gate.
One container stack inside your VPC or Kubernetes cluster. Zero code storage, zero telemetry — and it keeps working with outbound internet fully severed.
The contrast: CodeRabbit gates self-hosting behind an Enterprise contract with a 500-seat floor (as of July 2026). Graphite, Ellipsis, and Copilot offer no self-host at all.
Point Gatekeep at any OpenAI-compatible endpoint you already trust — Bedrock, Azure OpenAI, or local vLLM/Ollama with open-weight models. Your inference, your keys, no mandatory subprocessor.
The contrast: CodeRabbit, Graphite, Ellipsis, Copilot, and Cursor BugBot are all locked to managed model pipelines. Gatekeep treats the model as yours to choose.
A blocking status check locks the merge button on policy violations. Overrides require a written justification, and every decision lands in an immutable, exportable audit log — evidence your SOC 2 / ISO 27001 program consumes.
The contrast: CodeRabbit's checks are advisory — bypassable with a checkbox, no justification, no immutable log. Gatekeep makes the gate enforceable and the override accountable.
Add the Gatekeep GitHub App, or run docker compose up inside your
own network. One container stack, no code leaves your perimeter.
$ docker compose up -d
✔ Network gatekeep_default Created
✔ Container gatekeep-gateway Started
✔ Container gatekeep-reviewer Started
gatekeep ready on :8080 · egress: DISABLED
One endpoint in config.yaml — Bedrock, Azure OpenAI, or a local
vLLM/Ollama instance. Gatekeep uses your inference for every review.
llm:
# OpenAI-compatible: vLLM, Ollama, Azure, Bedrock/LiteLLM
base_url: http://vllm.internal:8000/v1
model: qwen2.5-coder-32b-instruct
api_key_env: LLM_API_KEY # your vault, never inlinedYour rules: what blocks a merge, who can override, what a justification must contain. Enforced as a blocking status check, logged on every decision.
review:
gate_threshold: high # findings ≥ high lock the merge
policy_file: /config/policy.md
governance:
override_allowlist: ["alice", "bob"]
min_justification_len: 40 # no checkbox escape hatchThat's it. Open a PR and the gate is live — see the real output.
How Gatekeep's BYO-LLM, air-gap-capable merge gate compares. Competitor behavior reflects publicly documented terms as of July 2026.
| Capability | Gatekeep | CodeRabbit | Graphite / Ellipsis | Copilot / BugBot |
|---|---|---|---|---|
| True air-gapped self-host | Yes | Enterprise only, 500+ seats | No self-host | No self-host |
| Bring your own LLM | Yes | Self-hosted Enterprise only | Managed pipeline | Managed pipeline |
| Enforceable, non-bypassable gate | Yes | Advisory / checkbox bypass | Advisory | Advisory |
| Immutable override audit log | Yes | Records who, not why | None | None |
| Self-host entry price | $499/mo flat | Contact sales, 500+ seats | — | — |
Early-access pricing. Design-partner terms available now.
$0
Public repos and small trials. Opens up when the code goes public. Early access
See self-host docs$19 / user / mo
Billed annually ($24 month-to-month), minimum 3 seats. Managed AI review, undercutting CodeRabbit Pro.
Join the waitlist$499 / mo flat
Up to 25 devs ($25/dev/mo above 25) — ≈$875/mo for a 40-dev team. Regulated and IP-sensitive teams that need review inside their own network. Flat pricing, no enterprise sales cycle.
Talk to the teamfrom $99 / mo
Discounted, hands-on onboarding onto any deployment — cloud or self-hosted, up to 15 devs. Shape the roadmap with direct access to the founding team.
Join the design-partner programThe wedge: competitors treat self-host as an enterprise negotiation — CodeRabbit's self-hosted tier requires an Enterprise contract with 500+ user seats, contact sales, no published price (as of July 2026). Gatekeep makes self-host a flat $499/mo product, not a negotiation. Self-host without the enterprise sales cycle. Full side-by-sides: vs CodeRabbit · vs Qodo · vs Greptile · vs Kodus
Gatekeep stores zero source code and sends zero telemetry. In self-hosted mode, code and inference never leave your network — we have no servers in the path and no ability to see your code.
Yes. With a self-hosted deployment and a local model endpoint, Gatekeep runs with all outbound internet access severed. Nothing about a review requires a call to our infrastructure or any third-party API.
Any OpenAI-compatible endpoint: AWS Bedrock, Azure OpenAI, or self-hosted open-weight models via vLLM or Ollama (e.g. Qwen2.5-Coder, Llama-3-70B). You bring the model and the keys; we bring the review pipeline and the gate.
GitHub only, today. GitLab support is on the near-term roadmap. We'd rather do one integration properly than spread thin — if GitLab is a dealbreaker for your pilot, tell us and we'll factor it into sequencing.
Yes. Paste the ticket text or a linked issue/PRD reference into the PR description; Gatekeep injects it as intent context and reviews the change against your stated requirements — the same lightweight mechanism as the policy file, no third-party Jira integration required.
Gatekeep registers as a required, blocking status check in your branch-protection rules. If a PR violates your policy, the merge stays locked. An authorized override requires a written justification, and the actor, target, timestamp, and reason are written to an immutable, exportable audit log.
Not silently. There is no "ignore checks" checkbox. An override is possible only with a recorded written justification — which is exactly the evidence auditors want to see.
Gatekeep produces the audit evidence those frameworks ask for — immutable override logs, enforced controls, and a zero-egress deployment model. We are not ourselves SOC 2 certified yet and won't claim to be. Compliance is a property of your whole program, not one tool, but Gatekeep is built to be the reviewable, evidence-generating control in it.
A paid pilot from $99/mo with direct access to the founding team, a hands-on install, and real influence over the roadmap. In exchange we ask for candid feedback and, where you're willing, a reference. Short, no long lock-in.
Gatekeep is onboarding design partners now. If your team ships fast and can't send code to a SaaS reviewer, you're exactly who we built this for.
db/query.pyRequest input is concatenated into a raw SQL string. Use a parameterized query.