Claude Code · Codex · Gemini CLI · Copilot

Engineering governance,
built into every commit.

Manta reviews every change on your own machine, blocks what must not land, and produces audit-ready evidence of the decision.

Three agents review each commit. On push, Manta routes the change to up to nine relevant agents. It runs on your own AI account — we never see your code, and we never bill you for tokens.

payments-api — git push origin main
$ git push origin main

  Manta Pre-Push Review
  12 files changed · routed to 8 of 9 agents

  security-sentinel        CRITICAL 0  WARNING 1
  compliance-guardian      CRITICAL 0  WARNING 1
  spec-guardian            CRITICAL 0  WARNING 0
  zero-trust-guardian      CRITICAL 0  WARNING 0
  observability-guardian   CRITICAL 0  WARNING 0
  test-architect           CRITICAL 0  WARNING 3
  perf-analyzer            CRITICAL 0  WARNING 0
  code-quality             CRITICAL 0  WARNING 0

  CONFIRMED  src/api/users.ts:84
  Email address written to request log · GDPR Art. 5(1)(c)
  Raised independently by security-sentinel and compliance-guardian.

  ✗ PUSH BLOCKED · 1 confirmed warning
  Single-agent warnings are advisory and did not block.
  Recorded to the enforcement trail.

The data boundary

Your source never leaves your machine

Reviews execute locally, on your runner, with your key. What reaches the control plane is findings metadata — never a diff, never a snippet, and never the scanner's raw matches, because those lines are the credentials it just caught.

What we receive

{
  "severity": "CRITICAL",
  "path": "src/auth/login.js",
  "line": 12,
  "rule": "hardcoded-credential",
  "cwe": ["CWE-798"],
  "verdict": "BLOCK"
}

What we never receive

  • Diffs, hunks, file contents
  • Code snippets — including the offending line
  • Raw scanner matches (they contain the secret)
  • Commit messages
  • Environment variables or credentials files

You do not have to take our word for it. Run manta-emit.sh --dry-run and read the exact bytes, or --explain for the redaction policy. Redaction fails shut: if the policy cannot load, nothing is emitted at all.

The pipeline

Reviewed where it is cheapest to fix

The split is not arbitrary. Three agents run at commit — the checks where committing is itself the harm. The rest run at push, routed by what the change actually touches, so a CSS commit does not wake the migration guardian.

1 · Install

One script. Agents, git hooks and config land in your repo. No service to run.

2 · Commit

Three agents read the staged diff — secrets, quality, migrations. Criticals block.

3 · Push

Up to nine agents, selected by what changed. A warning blocks only when two agents agree.

4 · Evidence

Metadata syncs in the background — never on the commit path — and becomes the trail.

A dense review takes minutes, not seconds. It is reading your code, not linting it. Manta is never in the path of the AI call: reviews bill to your own provider account at cost, with no inference markup.

Free edition and Enterprise

13 more agents, and a record of what they decided

Manta's free edition is 20 agents and 21 commands, open source under Apache 2.0, and it stays that way. Enterprise is 33 and 32 — the difference is the governance layer, and eleven of those agents are served from the control plane while your subscription is active.

Free forever

20 agents · 21 commands · Apache 2.0 · no account

security-sentinelcode-qualitydb-migration-guardianperf-analyzercode-writerdoc-keeper+ 14 more

github.com/mantacron/manta →

Enterprise

33 agents · 32 commands · hosted enforcement record · 11 agents served under subscription

compliance-guardianzero-trust-guardianspec-guardianobservability-guardiantest-architectpentest-agententerprise-readiness-agentaudit-reporterlog-analystproject-risk-analystproject-architect

Talk to us →

 MantaManta Enterprise
Agents2033
Commands2132
Review at git commit3 agents3 agents
Review at git push3–4 agents, trigger-routedup to 9 agents, trigger-routed
A warning at pushblocksblocks only when two agents agree
Code generation — /write, /scaffold, /ui
Research → Plan → Implement, with a GO/NO-GO gate
Security scanOWASP + secrets+ dependency CVEs, licence conflicts, dead deps
Spec governance — spec-guardian
Compliance — GDPR, HIPAA, PCI-DSS, SOC 2
Zero-trust — IAM, mTLS, RBAC, token TTLs
Observability checks on new endpoints
Test-coverage enforcement
Health scoring and trends across audits
Penetration testing — /penteston demand
Log analysis — /log-analysis
Hosted enforcement record and evidence bundles
LicenceApache 2.0, free foreverCommercial subscription

Enterprise adds 13 agents; 11 of them are served from the control plane while a subscription is active. Counts are generated from the shipped rosters, not typed by hand.

Compliance enforcement

GDPR, HIPAA, PCI-DSS and SOC 2 checks run on every push — not the week before the audit. Active only for the frameworks your CONSTITUTION.md declares.

Risk & governance

Health scores, trend tracking against previous reports, and risks surfaced before the architecture sets — real visibility, not last quarter's gut feeling.

Security depth

OWASP Top 10 and hardcoded secrets on the commit path; zero-trust and observability checks at push; dependency CVEs and licence conflicts with /security-scan; penetration testing on demand with /pentest.

Spec & architecture

Architectural drift, scope violations and API contract deviations are caught before the PR is opened. The spec stays the single source of truth.

Compliance evidence

Evidence beyond the dashboard

GRC platforms integrate at the cloud layer and cannot see the code layer. Manta sits exactly there, so it can evidence controls they can only ask you about.

Q3 2026 · GDPR · acme

Art. 32 — security of processing1,284 commits screened
Art. 17 — right to erasure1 gap, 2 repositories
Change management4 criticals blocked at commit
Enforcement gaps7 bypasses, each with actor

Bundles lead with their own limitations — uncovered repositories, unresolved checks, every bypass, and every verdict replayed from an earlier identical review rather than freshly run — because a document that overstates coverage is worse than none.

Every bypass, with its reason

A gate nobody can override is a gate people route around. Overrides are allowed and recorded — who, when, which finding, and why.

Digests, without opening anything

A periodic summary to Slack or any JSON endpoint. Coverage, blocks and bypasses as totals — never a leaderboard of your engineers.

The same gate in CI

The review that runs on a laptop runs on a pull request, so a machine without Manta installed cannot quietly skip it.

Built for

The teams that cannot afford to get this wrong.

  • Engineering teams in regulated industries — fintech, healthcare, legal, defence
  • Organisations subject to GDPR, HIPAA, PCI-DSS or SOC 2 audits
  • CTOs and VPs who need codebase health metrics for leadership and board reporting
  • Teams scaling past ten engineers, where manual review becomes the bottleneck
  • Security-conscious organisations that want continuous coverage between pentests
  • Teams whose code may not leave the building — reviews run locally, and only findings metadata is ever sent
3 at commit · up to 9 at push
Agents per review, routed by what the change touches
The commit gate stays fast because it is deliberately small.
0 lines of code
Leave your machine — findings metadata only
Verify it yourself with manta-emit.sh --dry-run.
No inference markup
Fixed subscription; AI usage billed by your own provider
We never proxy a token.

Get in touch

Let's talk about your team.

Tell us a little about your situation and we will get back to you within one business day. No sales pitch — a direct conversation about whether Manta fits.

Opens your mail client — nothing is submitted to a server.

Prefer email? ray@mantacron.com