Open source · Built on Claude

Twenty analyzers.
One gate.
Claude reads the verdict.

Noeticar runs ruff, mypy, bandit, semgrep, trivy and 15 more analyzers, merges them into a single 0–100 quality score, then asks Claude to review the result like a senior engineer: what's real, what's noise, and what the tools missed.

View on GitHub How Claude is used

# scan + Claude second opinion
$ noeticar scan . --second-opinion
  ruff ✓  mypy ✓  bandit ✓  semgrep ✓  radon ✓  +15
  score 78/100   gate FAIL (threshold 85)
  41 findings · 3 high · 12 medium

claude-sonnet-5-5 · second opinion
  agree     B608 SQL built with f-string  api/search.py:88
  dispute   B105 "token" is a constant name, not a secret
  missed    retry loop never backs off  queue.py:142
  reorder   fix B608 first — it is user-reachable
  confidence 0.82 · 31.4k in / 1.9k out tokens

20static-analysis runners

44MCP tools for Claude Code

1,700+unit tests

MITopen source, self-hosted

How it works

Analyzers find candidates. Claude decides what matters.

Static analysis is precise but noisy; LLM review is flexible but needs grounding. Noeticar gives Claude the structured scan, code snippets and your team's knowledge base, and asks for a verdict it can defend.

01 · SCAN

Run every analyzer once

UCC orchestrates 20 runners in parallel, normalises severities and produces one score, SARIF and an SBOM.

02 · REVIEW

Claude second opinion

Claude Sonnet 5.5 agrees, disputes, adds missed issues and re-ranks priorities, returned as strict JSON with a confidence value.

03 · LEARN

Remember the decisions

GloryStory, a local knowledge base, stores accepted and rejected findings so the next review starts from your team's history.

Built on Claude

Why Claude does the reviewing

Noeticar is built on Claude by Anthropic. The Claude API is the default provider, and the MCP server is designed for Claude Code.

Long context. A whole scan, the relevant snippets and prior KB decisions fit in a single request, so the review sees the full picture.
Reasoning you can audit. Every dispute carries a reason and a suggested severity; nothing is silently dropped.
Cost routing. The Plano router sends cheap triage to Haiku 5.5 and deep reviews to Opus 5.5, based on hotspot, impact and diff-size signals.

AgreeB608 Query string built with an f-string reaches cursor.execute from a request parameter.

DisputeB105 TOKEN_HEADER = "X-Token" is a header name, not a credential. Suggest: info.

MissedRetry loop in queue.py has no backoff or cap; under an outage it hammers the broker.

ReorderFix the SQL issue before the 9 typing errors: it's the only finding reachable from user input.

Model routing

The right Claude model for each job

Triage

claude-haiku-5-5

Small diffs and low-impact changes: fast, cheap first pass.

Review

claude-sonnet-5-5

Default second opinion for every scan.

Deep dive

claude-opus-5-5

Hotspot files and high blast-radius changes.

MCP server

44 tools your Claude Code agent can call

Connect once with claude mcp add noeticar -- noeticar-mcp. Your agent can scan before committing, check the gate, look up past decisions and ask for a second opinion without leaving the session.

noeticar_scan
noeticar_gate
noeticar_fix
noeticar_dimensions
noeticar_sbom
noeticar_second_opinion
noeticar_analyze
noeticar_hotspots
noeticar_impact
noeticar_dora_metrics
noeticar_health
noeticar_workspace_health
noeticar_kb_search
noeticar_kb_add
noeticar_kb_link
noeticar_scan_and_learn
noeticar_scan_changed_and_explain
noeticar_unified_search
noeticar_semantic_search
noeticar_plano_route
noeticar_plano_budget
noeticar_plano_guardrails
noeticar_baseline
noeticar_policy_packs
… and 20 more

Questions

FAQ

Which AI model does Noeticar use?

Noeticar is built on Claude by Anthropic. The second-opinion review defaults to Claude Sonnet 5.5 through the Claude API, and the router can use Claude Haiku 5.5 and Claude Opus 5.5. Local models through Ollama are supported for air-gapped setups.

Does my code leave my machine?

The analyzers run locally. Only the scan summary and the snippets you allow in [second_opinion.scope] are sent to the Claude API, and you can cap the context size.

Does Claude change my code?

No. Claude produces a review; fixes come from deterministic auto-fixers (e.g. ruff) or from your own agent, and the quality gate stays the source of truth.

Is it free?

Yes, MIT-licensed and self-hosted. You bring your own Anthropic API key for the Claude review.

Twenty analyzers.One gate.Claude reads the verdict.