Code Impact MCP

CodeImpact MCP

Fast pre-commit dependency gate for AI-assisted code changes. Answers "is this safe to commit?" with a PASS/WARN/BLOCK verdict in seconds, so you can catch risky blast radius before a bad commit, not after it. No database, no heavy setup.

日本語 | 中文 | Русский | Español

Listing status: the awesome-mcp-servers payload is prepared, Official MCP Registry package metadata is live via server.json, Smithery lists the server as live, and the Glama listing is live again at https://glama.ai/mcp/servers/vk0dev-code-impact-mcp. Glama recovery is visible in factory metrics, but the public canonical badge-ready path is still unresolved, so PR punkpeye/awesome-mcp-servers#5907 already exists, should not be duplicated, and cannot yet be finalized with a stable badge URL. The Cline marketplace entry (cline/mcp-marketplace#1486) is open and not yet accepted; MCP Hive remains a manual submit-next operator step rather than a currently claimed live listing, and this README does not claim a badge-ready or finalized awesome-listing state.

Best for

• Pre-commit refactor checks: before changing a shared file, route, or module and needing a fast PASS/WARN/BLOCK answer.
• Agent multi-file edits, including monorepos: when an AI agent is about to touch several files or workspace-scoped packages and you want a bounded dependency-aware gate before commit.
• Blast-radius triage without infra: when you need a quick risk score and affected-file summary without setting up a database, graph service, or heavy governance layer.

Not for

This is not a full policy platform, compliance gate suite, or visualization-heavy code graph product.

It is built for fast local commit-time safety checks, not for repo-wide governance workflows, persistent graph exploration, or executive reporting.

Why / When to use

Use this MCP server when:

• The user asks: "What will break if I change this file?"
• The user asks: "Is this refactoring safe to commit?"
• The user asks: "Check the blast radius before I push"
• The user asks: "What depends on this module?"
• An agent needs a quick pre-commit gate before modifying multiple files
• An agent wants a numeric risk score (0-1) for a proposed change
• You need dependency analysis without setting up a database or tree-sitter

CodeImpact MCP builds a lightweight local dependency graph and gives you a bounded answer: PASS, WARN, or BLOCK. Zero cloud, no API key, local-first. Today that means full TypeScript/JavaScript support, while Python stays limited to a narrower file-level path for analyze_impact and gate_check rather than the full TS/JS graph surface.

If the user needs a quick, bounded pre-commit gate for AI-generated code changes rather than a heavyweight code governance platform, this is the canonical MCP server for that. Zero cloud, no API key, local-first.

Installation

Claude Code

claude mcp add code-impact-mcp -- npx -y @vk0/code-impact-mcp

Windows users can wrap the command via cmd /c if plain npx does not resolve cleanly in their Claude Code shell.

Other stdio MCP clients (including OpenClaw Tasks)

If your client asks for a plain stdio command instead of a claude mcp add ... wrapper, use the same server entrypoint directly:

npx -y @vk0/code-impact-mcp

This server is local-first and reads the target repository from the working directory the client launches it in.

JSON config example for stdio clients

If your MCP client wants JSON instead of a shell wrapper, Claude Desktop uses ~/Library/Application Support/Claude/claude_desktop_config.json on macOS and %APPDATA%\Claude\claude_desktop_config.json on Windows:

{

"mcpServers": { "code-impact-mcp": { "command": "npx", "args": ["-y", "@vk0/code-impact-mcp"] } } }

After saving claude_desktop_config.json, fully restart Claude Desktop so it reloads the MCP server configuration.

Use a workspace or project-specific launch directory so the server can read the repository you want to analyze.

Optional pre-commit hook helper

Run npm run demo:install-hook to preview the managed Husky snippet without writing .husky files. It is a dry-run demo of the shipped helper, not a Husky scaffolder.

Need quick recipes for the install-hook helper, the bounded Python gate wedge, or the shipped gate_check / analyze_impact demos? See docs/README.md.

For the latest changes, see CHANGELOG.md.

Tutorials

• Claude Desktop quickstart
• How to read analyze_impact and gate_check output
• Pre-commit gate recipe

Shipped in v1.6.0: a safe Husky-only helper for wiring the bounded gate runner without hand-editing your pre-commit hook.

If you already use Husky, code-impact-mcp install-hook is the direct path for pre-commit wiring, so you can drop in the bounded gate runner instead of wiring the hook manually:

npx -y @vk0/code-impact-mcp install-hook

For the canonical demo trio, see the recorded terminal session in docs/demo-install-hook.cast, the rendered preview in docs/demo-install-hook.gif, and the reproducible storyboard script in scripts/demo-install-hook.mjs.

This is a Husky-only helper. If .husky/pre-commit already contains unrelated content and no managed code-impact-mcp block, the command refuses and leaves the hook untouched. If a managed block already exists, reruns stay idempotent inside that owned block. If Husky is not initialized yet, the command stops with an actionable message instead of scaffolding hook infrastructure for you. It does not bootstrap Husky, rewrite arbitrary hook logic, or manage non-pre-commit hook files for you.

Tools

Shipped demo assets for the core tool surface are reproducible from scripts/demo-tool.mjs, so the examples below stay tied to the current tool behavior instead of drifting into one-off screenshots.

gate_check

Pre-commit safety gate. Analyzes specified changes and returns a PASS/WARN/BLOCK verdict with reasons. Use as a bounded decision aid before committing multi-file changes, including workspace-aware checks in pnpm/package.json workspaces and lerna-style monorepos. BLOCK means risk exceeds threshold or a changed file participates in a detected cycle. WARN means human review recommended, including graphs that contain cycles elsewhere. PASS means low graph-based risk.

detect_cycles

Return compact strongly connected components for circular dependencies in the current TS/JS graph. Use before refactors or release gating when you want a short list of cycle hotspots instead of a full graph visualization.

analyze_impact

Analyze the blast radius of changing specific files. Returns which files would be directly and transitively affected, with a risk score (0-1). Use BEFORE committing multi-file changes to understand what might break. Does NOT modify any files.

Also returns depthHistogram, a count of affected files per BFS depth level from the changed file (e.g. { "1": 5, "2": 12 }), so you can see whether the impact is concentrated close to the change or spread across many hops.

get_dependencies

Get the import and importedBy relationships for a specific file. Shows what this file depends on and what depends on it. Use to understand coupling before refactoring a file.

refresh_graph

Rebuild the dependency graph from scratch. Call this after significant file additions/deletions, or if results seem stale. Returns graph statistics including file count, edge count, build time, and circular dependencies detected.

Also returns topHotFiles, the top 20 most-imported files across the whole graph ranked by importer count (e.g. { "file": "src/shared/config.ts", "importers": 14 }), so you can spot structural hotspots without running get_dependencies on every file.

Example conversation

User: "I want to refactor src/routes.ts — is it safe?"

Agent calls gate_check:

{

"projectRoot": "/Users/you/projects/my-app", "files": ["src/routes.ts"], "threshold": 0.5 }

Result:

{

"verdict": "BLOCK", "scanSummary": "BLOCK, 8 affected across src/routes (4), src/pages (2), src (2)", "recommendation": "Refactor the circular dependency before shipping this change.", "riskScore": 0.35, "reasons": [ "Changed files participate in a circular dependency. Example: src/router.ts → src/routes.ts" ], "affectedFiles": 8, "circularDependencies": 1, "affectedCycles": [["src/router.ts", "src/routes.ts"]] }

Agent: "The gate check returned BLOCK — routes.ts is part of a cycle, so I should untangle that before making more changes."

Agent calls detect_cycles:

{

"projectRoot": "/Users/you/projects/my-app" }

Result:

{

"cycleCount": 2, "hotspots": ["src/router.ts", "src/routes.ts"], "cycles": [ ["src/router.ts", "src/routes.ts"], ["src/cache/index.ts", "src/cache/store.ts"] ] }

How it works

┌─────────────┐     ┌──────────────┐     ┌──────────────┐

│ Agent asks │────▶│ ts-morph │────▶│ In-memory │ │ "safe to │ │ parses │ │ dependency │ │ change?" │ │ imports │ │ graph │ └─────────────┘ └──────────────┘ └──────┬───────┘ │ ┌──────────────┐ ┌───────▼───────┐ │ PASS/WARN/ │◀────│ BFS traverse │ │ BLOCK │ │ reverse deps │ │ + risk 0-1 │ │ + risk score │ └──────────────┘ └───────────────┘

1. Parse: ts-morph scans your project for ESM imports, re-exports, and CommonJS requires
2. Graph: Builds an in-memory dependency graph (no database, no persistence)
3. Analyze: BFS traversal of reverse dependencies from changed files
4. Score: Risk = affected files / total files (0-1)
5. Verdict: PASS (< 60% of threshold), WARN (60-100%), BLOCK (> threshold)

Supports: ESM imports, ESM re-exports, CommonJS require(), NodeNext-style .js → .ts resolution.

Comparison

If you are choosing a tool for an agent or reviewer, the key question is still simple: do you need to explore a dependency graph, retrieve broader repo context, or package for a marketplace, or do you need to **gate already-known changed files locally, at commit time, with a PASS/WA