The static analyzer your AI agent reads — fix-ready, machine-readable scan reports over MCP.
io.github.abhinavteja123/codemore MCP Server
CodeMore provides a static analyzer for AI agents, producing fix-ready, machine-readable scan reports over MCP. The server scans code and returns findings designed for the agent that authored the code, including the exact fix and verification criteria so the agent can close its own findings.
🛠️ Key Features
Static analyzer that generates machine-readable scan reports over MCP
Fix-ready output including exact fixes
Verification criteria included with each finding
Readme references: Rule Catalog (docs/rules) and Report Schema (docs/schema.md)
🚀 Use Cases
Integrating SAST/static analysis into AI agent workflows
Supporting developer tools that run scans and feed structured results to agents
Using with TypeScript-oriented development and related environments (topics listed)
AI agents ship code fast — and ship bugs fast. CodeMore scans the code, then hands the agent that wrote it a machine-readable report with the exact fix and the criteria to verify it. The agent closes its own findings.
AI-assisted coding ships vulnerabilities at a measured, growing rate:
Finding
Source
45% of AI-generated code carries an OWASP Top-10 vulnerability
Veracode 2025/26
98% of 1,072 scanned vibe-coded sites had ≥ 1 security flaw
Symbiotic
70% of audited Lovable apps shipped with Supabase RLS disabled
DEV
2× baseline secret-leak rate on AI-tool-assisted commits
GitGuardian SOSS 2026
35 CVEs/month attributed to AI-generated code (was 6/month in January)
March 2026
Existing scanners (SonarQube, DeepSource, Snyk) target human reviewers sitting at dashboards. But this code wasn't written by a human — and the LLM that wrote it is fully capable of fixing its own bug, if the report is shaped for a machine reader.
That's the wedge. CodeMore is not another SAST dashboard. It's the report contract between a scanner and a coding agent: every finding carries a suggestedFix with a patch template and explicit verificationCriteria — not just "here's a problem," but "here's exactly how to know you fixed it."
The agent that wrote the bug can also write the fix — if it can read the report.
Quick start
CLI
bash
npx codemore@latest scan .
Prints a summary to the terminal. Add --json for the full report on stdout, or --out codemore-report.json to write it to disk: every finding pinned to file:line:column with rule citation, fix template, and verification criteria. Pipe it to your agent and watch findings close.
bash
npm install -g codemore # once — or prefix each command below with `npx codemore@latest`
codemore update # global install pinned an old version? this pulls latest
codemore scan . --fail-on BLOCKER # CI gate: non-zero exit on any BLOCKER
codemore scan . --external-tools ruff,biome # opt in to external tools
codemore scan . --external-tools all # ruff · golangci-lint · clippy · biome · bandit · gitleaks · npm-audit · pip-audit
codemore scan . --format sarif --out codemore.sarif # GitHub code scanning (upload-sarif)
codemore fix . --rule <id> --write # agentic fix loop from the CLI (needs an LLM API key)
codemore baseline create # adopt on an existing repo without drowning in legacy findings
MCP server — Cursor, Claude Code, Codex, Claude Desktop
bash
npx codemore mcp # print config snippet + every client's config path
npx codemore mcp install --client cursor # merge into existing config (backs up first, --dry-run supported)
PR-comment bot; only fails the build on findings new since the committed .codemore-baseline.json. A full copy-paste security gate chaining CodeMore + Ruff + Biome + Bandit + Gitleaks + npm-audit + pip-audit + Checkov lives at templates/.github/workflows/codemore-security-gate.yml (walkthrough).
Web Scanner (hosted)
Sign in at codemore.tech, paste a public GitHub URL or upload a ZIP — same report, same fingerprint, zero install.
What it catches
64 native rules across 6 packs, every rule mapped to a cited real-world incident class — not a hypothetical:
RLS-off, RLS-permissive (USING (true)), anon-key bundled to client
8 external adapters (off by default, opt in via --external-tools): ruff · golangci-lint · clippy · biome · bandit · gitleaks · npm-audit · pip-audit. Findings are namespaced ext:<tool>:<rule-id> — no collision with native rules; a missing binary skips silently instead of crashing.
The walker catches what .gitignore hides. When a developer "hides" a leaked secret file by gitignoring it, most scanners stop seeing it — but it's still on disk, in tarballs, in Docker images. CodeMore always scans secret-shaped filenames (.env*, *.pem, *.key, firebase-adminsdk*.json, credentials.json, .npmrc, .pypirc, …) even when gitignored. This is exactly how the audits found real production OpenAI keys, Google API keys, and Firebase admin SDK creds that other tools silently missed. Opt out with --respect-gitignore-fully.
Per-rule documentation: docs/rules — 64 pages, one per rule.
The report is the product
Every surface emits the same schema-stable codemore-report.json — verified byte-identical (modulo timestamps and instance IDs) on every release by test/parity.test.ts:
One brain, four skins. The schema is the API; surfaces are interchangeable.
Full report example — codemore-report.json v1.0.0
jsonc
{"schemaVersion":"1.0.0","tool":{"name":"codemore","version":"0.3.0"},"project":{"root":".","framework":"next.js","language":"typescript","fingerprint":"sha256:7f95f2c62e0d3ecea6f23…"},"summary":{"issuesTotal":42,"bySeverity":{"BLOCKER":2,"CRITICAL":5,"MAJOR":15,"MINOR":18,"INFO":2},"byCategory":{"security":12,"bug":7,"…":"…"},"filesAnalyzed":87,"linesOfCode":12450,"technicalDebtMinutes":1840},"issues":[{"id":"vibe-supabase-rls-disabled","ruleVersion":"1.2.0","instanceId":"01HZ…","severity":"BLOCKER","confidence":0.95,"category":"security","title":"Supabase table has no RLS policy","evidence":{"file":"supabase/migrations/001_init.sql","line":14,"column":1,"endLine":14,"endColumn":60,"snippet":"create table profiles (id uuid primary key, …);","matchedPattern":"create-table-without-rls"},"whyItMatters":"Public Supabase client can read/write all rows. 70 % of Lovable apps leak data through this.","citation":"https://codemore.tech/rules/vibe-supabase-rls-disabled","suggestedFix":{"type":"code-patch","instructions":"Add `ALTER TABLE profiles ENABLE ROW LEVEL SECURITY;` plus at least one policy scoped to authenticated users.","patchTemplate":"…","verificationCriteria":["Migration contains ALTER TABLE … ENABLE ROW LEVEL SECURITY","At least one CREATE POLICY exists for the table","Re-scan no longer reports vibe-supabase-rls-disabled for this file"]},"suppression":{"available":true,"directive":"// codemore-ignore: vibe-supabase-rls-disabled","scope":"same-line | next-line | file"}}],"agentInstructions":{"preamble":"You are fixing issues found by CodeMore. Apply patches one issue at a time. After each, request re-scan via validate_fix.","orderingHint":"blockers → criticals → majors","doNotTouch":["node_modules/**","*.lock",".env*"],"stopOn":"first-validator-failure"},"meta":{"rulesEnabled":64,"packsLoaded":["core-security","core-quality","vibe-supabase","…"],"scanDurationMs":4321}}
Schema source-of-truth: shared/report/schema.json. Breaking changes bump schemaVersion major and ship a migration guide.
Agentic fix loop
apply_fix runs a four-stage loop, up to 3 retries per finding — it terminates on first PASS and never silently keeps a failing patch:
validatorHarness.ts — applies the patch in a tempdir copy, re-runs the rule, re-runs file-scoped tests, returns pass | fail + diagnostics
LLM provider plug-ins: OpenAI · Anthropic · Gemini · local, configurable via workspace setting (VS Code) or CODEMORE_LLM_PROVIDER env (CLI)
Accuracy, measured honestly
Synthetic benchmarks lie; real codebases don't. Every release is audited against real projects, and the numbers are published — including the bad ones.
2026-07-07 audit (full report) — 7 codebases (Python app, 1.35M-LOC TS monorepo, polyglot Rust, synthetic ground-truth apps, self-scan): 100% of planted vulnerabilities detected, a real OpenAI key caught in a production .env, ~90% BLOCKER true-positive rate, self-scan noise reduced 85% after fixing four false-positive classes.
2026-06-12 audit (full report) — 10 real codebases, aggregate ~85% BLOCKER TP rate, above DeepSource's ≥ 75% production bar:
Per-project results
Project
Findings
BLOCKERs
TP rate
Notes
EchoVault
122
10
100%
Real Supabase RLS holes
ProofSnap
134
4
100%
AImentor
224
5
100%
Real OpenAI keys hidden by .gitignore
Hackathonnn
129
3
100%
shopsec
113
5
100%
Senti
155
8
100%
open-design
8,615
101
~80%
Gen ai
63
4
75%
Real Firebase admin SDK creds
codemore self
282
6
17%
Intentional landing-demo data
Aggregate (excl. self)
9,755
140
~85%
Quality gates, enforced in CI
100% TP / 100% FP on the 128-fixture corpus regression suite — every rule ships with at least one true-positive and one false-positive fixture under corpus/rules/<rule-id>/{tp,fp}/
Lifecycle gating: rules are experimental → beta → stable, promotion requires fixture pairs plus real-world FP-rate evidence via opt-in telemetry; rules below the precision bar are gated behind --enable-experimental or ship with reduced confidence so agents weight them lower
Lifecycle
Default
Promotion bar
experimental
off by default
one fixture pair
beta
on by default
≥ 3 fixture pairs + 14-day FP rate < 15%
stable
ships in default pack
30-day FP rate < 5% + reference apps clean
deprecated
warns, removed next major
—
What CodeMore does not catch
Out of scope by design — static analysis can't judge these, and an agent can't auto-fix them from a source diff. The rule is "agent-actionable or it's not a rule":
Full list with reasoning: docs/limitations.md. For these, pair CodeMore with OWASP ZAP, Burp Suite, checkov, or your IdP's compliance dashboard.
Telemetry — opt-in only
Off by default. Enable per-scan with --telemetry. Collected: tool version, hashed project fingerprint, surface, rule-fire events. Never collected: file paths, contents, snippets, evidence text — the endpoint enforces a Zod strict() schema and rejects any payload containing those keys with HTTP 400. 64 KB payload cap, per-fingerprint rate limiting, RLS denies all reads.
One brain (shared/), four skins (CLI, MCP, extension, Action), one report schema — kept honest by the parity test, not by a slogan.
Development
bash
git clone https://github.com/abhinavteja123/codemore
cd codemore
npm ci # postinstall skips binary downloads in dev automatically
npx tsc -p tsconfig.publish.json # type-check the publishable surface
npm run test:unit # unit tests (mocha + ts-node)
npx mocha --require ts-node/register test/parity.test.ts # surface-parity proof
node cli.js scan corpus/rules/vibe-no-rate-limit/tp --json --enable-experimental
cd web && npm ci && npm run dev # landing + dashboard + docs → localhost:3000
VS Code extension: F5 in VS Code, or npm run watch + code --extensionDevelopmentPath=.
Contributing
Two paths:
New rules — read CONTRIBUTING-RULES.md. The PR bot gates every submission: rule module + TP fixture (must fire) + FP fixture (must not fire) + docs page + pack registration. The bot is the first reviewer; humans review after it passes.
