Analyze Codebase for MCP
Scan a codebase to discover functions, REST endpoints, CLI commands, and data access patterns that are good candidates for MCP tool exposure, then produce a structured tool specification document.
When to Use
- Planning an MCP server for an existing project and need to know what to expose
- Auditing a codebase before wrapping it as an AI-accessible tool surface
- Comparing what a codebase can do versus what is already exposed via MCP
- Generating a tool specification document to hand off to
scaffold-mcp-server - Evaluating whether a third-party library is worth wrapping as MCP tools
Inputs
- Required: Path to the codebase root directory
- Required: Target language(s) of the codebase (e.g., TypeScript, Python, R, Go)
- Optional: Existing MCP server code to compare against (gap analysis)
- Optional: Domain focus (e.g., "data analysis", "file operations", "API integration")
- Optional: Maximum number of tools to recommend (default: 20)
Procedure
Step 1: Scan Codebase Structure
1.1. Use Glob to map the directory tree, focusing on source directories:
src/**/*.{ts,js,py,R,go,rs}for source files**/routes/**,**/api/**,**/controllers/**for endpoint definitions**/cli/**,**/commands/**for CLI entry points**/package.json,**/setup.py,**/DESCRIPTIONfor dependency metadata
1.2. Categorize files by role:
- Entry points: main files, route handlers, CLI commands
- Core logic: business logic functions, algorithms, data transformers
- Data access: database queries, file I/O, API clients
- Utilities: helpers, formatters, validators
1.3. Count total files, lines of code, and exported symbols to gauge project size.
Expected: A categorized file inventory with role annotations.
On failure: If the codebase is too large (>10,000 files), narrow the scan to specific directories or modules using the domain focus input. If no source files are found, verify the root path and language parameters.
Step 2: Identify Exposed Functions and Endpoints
2.1. Use Grep to find exported functions and public APIs:
- TypeScript/JavaScript:
export (async )?function,export default,module.exports - Python: functions not prefixed with
_,@app.route,@router - R: functions listed in NAMESPACE or
#' @exportroxygen tags - Go: capitalized function names (exported by convention)
2.2. For each candidate function, extract:
- Name: function or endpoint name
- Signature: parameters with types and defaults
- Return type: what the function produces
- Documentation: docstrings, JSDoc, roxygen, godoc
- Location: file path and line number
2.3. For REST APIs, additionally extract:
- HTTP method and route pattern
- Request body schema
- Response shape
- Authentication requirements
2.4. Build a candidate list sorted by potential utility (public, documented, well-typed functions first).
Expected: A list of 20-100 candidate functions/endpoints with extracted metadata.
On failure: If few candidates are found, broaden the search to include internal functions that could be made public. If documentation is sparse, flag this as a risk in the output.
Step 3: Evaluate MCP Suitability
3.1. For each candidate, assess against MCP tool criteria:
- Input contract clarity: Are parameters well-typed and documented? Can they be described in a JSON Schema?
- Output predictability: Does the function return structured data (JSON-serializable)? Is the return shape consistent?
- Side effects: Does the function modify state (files, database, external services)? Side effects must be clearly labeled.
- Idempotency: Is the operation safe to retry? Non-idempotent tools need explicit warnings.
- Execution time: Will it complete within a reasonable timeout (< 30 seconds)? Long-running operations need async patterns.
- Error handling: Does it throw structured errors or fail silently?
3.2. Score each candidate on a 1-5 scale:
- 5: Pure function, typed I/O, documented, fast, no side effects
- 4: Well-typed, documented, minor side effects (e.g., logging)
- 3: Reasonable I/O contract but needs wrapping (e.g., returns raw objects)
- 2: Significant side effects or unclear contract, needs substantial adaptation
- 1: Not suitable without major refactoring
3.3. Filter candidates to those scoring 3 or above. Flag score-2 items as "future candidates" requiring refactoring.
Expected: A scored and filtered candidate list with suitability rationale for each.
On failure: If most candidates score below 3, the codebase may need refactoring before MCP exposure. Document the gaps and recommend specific improvements (add types, extract pure functions, wrap side effects).
Step 4: Design Tool Specifications
4.1. For each selected candidate (score >= 3), draft a tool specification:
- name: tool_name
description: >
One-line description of what the tool does.
source_function: module.function_name
source_file: src/path/to/file.ts:42
parameters:
param_name:
type: string | number | boolean | object | array
description: What this parameter controls
required: true | false
default: value_if_optional
returns:
type: string | object | array
description: What the tool returns
side_effects:
- description of any side effect
estimated_latency: fast | medium | slow
suitability_score: 5
4.2. Group tools into logical categories (e.g., "Data Queries", "File Operations", "Analysis", "Configuration").
4.3. Identify dependencies between tools (e.g., "list_datasets" should be called before "query_dataset").
4.4. Determine if any tools need wrappers to:
- Simplify complex parameter objects into flat inputs
- Convert raw return values to structured text or JSON
- Add safety guards (e.g., read-only wrappers for database functions)
Expected: A complete YAML tool specification with categories, dependencies, and wrapper notes.
On failure: If tool specifications are ambiguous, revisit Step 2 to extract more detail from source code. If parameter types cannot be inferred, flag for manual review.
Step 5: Generate Tool Spec Document
5.1. Write the final specification document with these sections:
- Summary: Codebase overview, language, size, and analysis date
- Recommended Tools: Full specifications from Step 4, grouped by category
- Future Candidates: Score-2 items with refactoring recommendations
- Excluded Items: Score-1 items with exclusion rationale
- Dependencies: Tool dependency graph
- Implementation Notes: Wrapper requirements, authentication needs, transport recommendations
5.2. Save as mcp-tool-spec.yml (machine-readable) and optionally mcp-tool-spec.md (human-readable summary).
5.3. If an existing MCP server was provided, include a gap analysis section:
- Tools in the spec but not yet implemented
- Implemented tools not in the spec (possibly stale)
- Tools with specification drift (implementation diverges from spec)
Expected: A complete tool specification document ready for consumption by scaffold-mcp-server.
On failure: If the document exceeds reasonable size (>200 tools), split into modules with cross-references. If the codebase has no suitable candidates, produce a "readiness assessment" document with refactoring recommendations instead.
Validation
- All source files in the target codebase were scanned
- Candidate functions have extracted names, signatures, and return types
- Each candidate has a suitability score with written rationale
- Tool specifications include complete parameter schemas with types
- Side effects are explicitly documented for every tool
- The output document is valid YAML (parseable by any YAML library)
- Tool names follow MCP conventions (snake_case, descriptive, unique)
- Categories and dependencies form a coherent tool surface
- Gap analysis is included when an existing MCP server was provided
- Future candidates section lists refactoring steps needed for score-2 items
Common Pitfalls
- Exposing too many tools: AI assistants work best with 10-30 focused tools. Prioritize breadth of capability over depth. Resist exposing every public function.
- Ignoring side effects: A function that "just reads" but also writes to a log or cache still has side effects. Audit carefully with
Grepfor file writes, network calls, and database mutations. - Assuming type safety: Dynamic languages (Python, R, JavaScript) may have functions with no type annotations. Infer types from usage patterns and tests, but flag uncertainty in the spec.
- Missing authentication context: Functions that work in an authenticated web request may fail when called via MCP without session context. Check for implicit auth dependencies such as session cookies, JWT tokens, or environment-injected credentials.
- Over-engineering wrappers: If a function needs a 50-line wrapper to be MCP-compatible, it may not be a good candidate. Prefer functions that map naturally to tool interfaces.
- Neglecting error paths: MCP tools must return structured errors. Functions that throw untyped exceptions need error-handling wrappers.
- Conflating internal and external APIs: Internal helper functions called by other internal code are poor MCP candidates. Focus on functions designed for external consumption or clear boundary APIs.
- Skipping the gap analysis: If an existing MCP server is provided, always compare the spec against current implementation. Without gap analysis, you risk duplicating work or missing stale tools.
Related Skills
scaffold-mcp-server- use the output spec to generate a working MCP serverbuild-custom-mcp-server- manual server implementation referenceconfigure-mcp-server- connect the resulting server to Claude Code/Desktoptroubleshoot-mcp-connection- debug connectivity after deploying the serverreview-software-architecture- architecture review for tool surface designsecurity-audit-codebase- security audit before exposing functions externally