Skip to content

Meridian — DevOps Gate

A hard-blocking change gate for AI-generated code. Self-hosted, Apache-2.0, air-gap capable.

Meridian sits between a code change and your deploy. It analyses every diff through three gates, opens an RFC (Request for Change) that records the verdict, and blocks the change until that RFC is APPROVED or explicitly OVERRIDDEN. The decision is written to a tamper-evident (WORM) audit trail.

Unlike scanners that post advisory comments, Meridian is designed to stop the merge/deploy when a change is risky.

The problem

AI coding assistants (Claude, Cursor, Copilot, etc.) now generate a large share of the code that reaches production. They are fast and frequently wrong in subtle ways: a hardcoded secret, a child_process.exec with interpolated input, a dropped tenant filter, an eval() on request data.

Traditional tools react to this in three unsatisfying ways:

  • Linters / SAST (Semgrep OSS, SonarQube) — produce findings, but do not stop anything. The findings sit in a dashboard.
  • GHAS — cloud-only, per-seat pricing, and still advisory by default.
  • Code review — human, slow, and inconsistent at 2 a.m. when an agent opens 14 PRs.

None of them give you a single artifact that says "this exact diff was reviewed, here is the verdict, here is who overrode it and why" — and none of them refuse to let a bad change through.

How it works — the 3 gates

Every diff submitted to Meridian flows through three gates in order. The first gate to produce a blocking finding stops the pipeline.

flowchart LR
    A[Diff submitted] --> G1[Gate 1\nRisk Assessment\nregex pattern scan]
    G1 --> G2[Gate 2\nAST Architecture\nSemgrep rules arch-01..16]
    G2 --> G3[Gate 3\nLLM Review\nOllama -> DeepSeek -> Anthropic]
    G3 --> R{RFC verdict}
    R -->|clean| AP[APPROVED]
    R -->|risk found| BL[BLOCKED]
    BL -->|human decision| OV[OVERRIDDEN]
Gate What it does Engine
① Risk Assessment Fast regex scan for secrets, dangerous calls, known vuln patterns Built-in patterns + your MERIDIAN_RULES_PATH
② AST Architecture Check Structural rules (arch-01 … arch-16): missing auth, tenant-scope leaks, unsafe deserialization Semgrep-compatible .semgrep.yaml
③ LLM Review Reasoning pass that catches what regex/AST cannot, with a tiered, cost-capped router Ollama → DeepSeek → Anthropic

The outcome is an RFC that moves through a defined lifecycle:

DRAFT → BLOCKED | APPROVED | OVERRIDDEN → SUPERSEDED

See What is an RFC? for the full lifecycle.

Meridian vs the alternatives

Meridian Semgrep Team GitHub Advanced Security SonarQube Snyk
Blocks the deploy (not advisory) Yes No (advisory) No by default No No
RFC artifact per diff Yes No No No No
WORM audit trail Yes No No No No
LLM reasoning layer Yes No No No Partial
Self-hosted Yes Partial No (cloud) Yes No
Air-gap capable Yes No No Partial No
License Apache-2.0 Proprietary Proprietary LGPL/Commercial Proprietary
Indicative price $0 ~$450/mo $19/seat/mo $150+/mo Quote

Pricing figures are indicative public list prices and change frequently; treat them as order-of-magnitude. The point is not that the others are bad — it is that none of them block and none of them keep a WORM RFC.

Cost comparison

For a 10-developer team:

Tool Monthly Annual
GitHub Advanced Security ($19/seat) $190 $2,280
Semgrep Team ~$450 ~$5,400
SonarQube (paid tier) $150+ $1,800+
Meridian (Apache-2.0) $0 $0

Meridian's only running cost is your own infrastructure plus optional LLM API spend, which is hard-capped via LLM_DAILY_CAP_USD and can be $0 if you run Ollama only. See LLM cost control.

Quickstart

git clone https://github.com/weilmaschinchen/meridian.git
cd meridian
docker compose up -d --build

# Verify
curl -s http://localhost:3011/api/cra/health
# {"status":"ok", ...}

Full walkthrough: Docker Compose → Your first analysis.

Who is this for?

  • Solo developers shipping AI-generated code who want a safety net without a SaaS subscription → Solo dev scenario
  • Small teams that want a consistent, automated gate instead of relying on whoever reviews the PR → Small team scenario
  • Regulated industries (fintech, healthcare) that need a defensible, tamper-evident record of every change decision → Regulated scenario

Where to go next

Honesty first

This documentation calls out what Meridian cannot do yet (see Gaps and roadmap). For example, there is no MERIDIAN_AST_RULES_DIR — custom Semgrep rules currently require adding files to meridian/core/ast-spec/ or shipping a plugin. We would rather you know that up front.