codesurface MCP server

codesurface

MCP server that indexes your codebase's public API at startup and serves it via compact tool responses, saving tokens vs reading source files.

Parses source files, extracts public classes/methods/properties/fields/events, and serves them through 5 MCP tools. Works with Claude Code, Cursor, Windsurf, or any MCP-compatible AI tool.

Supported languages: C# (.cs), C++ headers (.h, .hpp, .hxx, .h++), Go (.go), Java (.java), Python (.py), TypeScript/JavaScript (.ts, .tsx, .js, .jsx)

Quick Start

Add to your .mcp.json:

{
  "mcpServers": {
    "codesurface": {
      "command": "uvx",
      "args": ["codesurface", "--project", "/path/to/your/src"]
    }
  }
}

Point --project at any directory containing supported source files (a Unity Assets/Scripts folder, a Spring Boot project, a .NET src/ tree, a Node.js/React project, a Python package, etc.). Languages are auto-detected.

Restart your AI tool and ask: "What methods does MyService have?"

CLAUDE.md Snippet

Add this to your project's CLAUDE.md (or equivalent instructions file). This step is important. Without it, the AI has the tools but won't know when to reach for them.

## Codebase API Lookup (codesurface MCP)

Use codesurface MCP tools BEFORE Grep, Glob, Read, or Task (subagents) for any class/method/field lookup. This applies to you AND any subagents you spawn.

| Tool | Use when | Example |
|------|----------|---------|
| `search` | Find APIs by keyword | `search("MergeService")` |
| `get_signature` | Need exact signature | `get_signature("TryMerge")` |
| `get_class` | See all members on a class | `get_class("BlastBoardModel")` |
| `get_stats` | Codebase overview | `get_stats()` |

Every result includes file path + line numbers. Use them for targeted reads:
- `File: Service.cs:32` → `Read("Service.cs", offset=32, limit=15)`
- `File: Converter.java:504-506` → `Read("Converter.java", offset=504, limit=10)`

Never read a full file when you have a line number. Only fall back to Grep/Read for implementation details (method bodies, control flow).

Tools

Tool	Purpose	Example
search	Find APIs by keyword	"MergeService", "BlastBoard", "GridCoord"
get_signature	Exact signature by name or FQN	"TryMerge", "CampGame.Services.IMergeService.TryMerge"
get_class	Full class reference card with all public members	"BlastBoardModel" → all methods/fields/properties
get_stats	Overview of indexed codebase	File count, record counts, namespace breakdown
reindex	Incremental index update (mtime-based)	Only re-parses changed/new/deleted files. Also runs automatically on query misses

search, get_signature, and get_class accept two optional filters:

• file_path: scope results to a directory prefix or exact file (e.g. "src/services/" or "src/services/MergeService.ts")
• include_tests: include test files in results (default false). Detects __tests__/, tests/, test/, *.test.*, *.spec.*, *_test.*, test_*

Tested On

Project	Language	Files	Records	Time
vscode	TypeScript	6,611	88,293	9.3s
Paper	Java	2,909	33,973	2.3s
client-go	Go	219	2,760	0.4s
langchain	Python	1,880	12,418	1.1s
pydantic	Python	365	9,648	0.3s
guava	Java	891	8,377	2.4s
immich	TypeScript	919	7,957	0.6s
fastapi	Python	881	5,713	0.5s
ant-design	TypeScript	2,947	5,452	0.9s
dify	TypeScript	4,903	5,038	1.9s
crawlee-python	Python	386	2,473	0.3s
flask	Python	63	872	<0.1s
cobra	Go	15	249	<0.1s
gin	Go	41	574	<0.1s
Unity game (private)	C#	129	1,018	0.1s

Line Numbers for Targeted Reads

Every record includes line_start and line_end (1-indexed). Multi-line declarations span the full signature:

[METHOD] com.google.common.base.Converter.from
  Signature: static Converter<A, B> from(Function<...> forward, Function<...> backward)
  File: Converter.java:504-506          ← multi-line signature

[METHOD] server.AlbumController.createAlbum
  Signature: createAlbum(@Auth() auth: AuthDto, @Body() dto: CreateAlbumDto)
  File: album.controller.ts:46          ← single-line

This lets AI agents do targeted reads instead of reading full files:

# Instead of reading the entire 600-line file:
Read("Converter.java")                     # 600 lines, ~12k tokens

# Read just the method + context:
Read("Converter.java", offset=504, limit=10)  # 10 lines, ~200 tokens

Benchmarks

Measured across 5 real-world projects in 5 languages, each using a 10-step cross-cutting research workflow.

Language	Project	Files	Records	MCP	Skilled	Naive	MCP vs Skilled
C#	Unity game	129	1,034	1,021	4,453	11,825	77% fewer
TypeScript	immich	694	8,344	1,451	4,500	14,550	68% fewer
Java	guava	891	8,377	1,851	4,200	26,700	56% fewer
Go	gin	38	534	1,791	2,770	15,300	35% fewer
Python	codesurface	9	40	753	2,000	10,400	62% fewer

Even with follow-up reads for implementation detail, the hybrid MCP + targeted Read approach uses 44% fewer tokens than a skilled Grep+Read agent and 87% fewer than a naive agent:

Per-question breakdown

See workflow-benchmark.md for the full step-by-step analysis across all languages.

Filtering What Gets Indexed

By default, codesurface skips common vendored, build, and VCS directories: node_modules, vendor, bin, obj, dist, build, target, .git, .venv, __pycache__, and a few dozen others. Git worktrees and submodules are also skipped.

To exclude additional paths:

Project-level (committed): create a .codesurfaceignore file at your project root with one glob per line.

generated/**
docs/**
**/*.pb.go

Per-instance (CLI): pass --exclude with comma-separated globs.

{
  "command": "uvx",
  "args": ["codesurface", "--project", "src", "--exclude", "generated/**,vendor/**"]
}

Other indexing flags:

• --include-submodules: index git submodules (skipped by default)
• --language <name>: pin to a single parser (e.g. --language cpp) instead of auto-detecting

Multiple Projects

Each --project flag indexes one directory. To index multiple codebases, run separate instances with different server names:

{
  "mcpServers": {
    "codesurface-backend": {
      "command": "uvx",
      "args": ["codesurface", "--project", "/path/to/backend/src"]
    },
    "codesurface-frontend": {
      "command": "uvx",
      "args": ["codesurface", "--project", "/path/to/frontend/src"]
    }
  }
}

Each instance gets its own in-memory index and tools. The AI agent sees both and can query across projects.

Setup Details

Using pip install:

pip install codesurface

{
  "mcpServers": {
    "codesurface": {
      "command": "codesurface",
      "args": ["--project", "/path/to/your/src"]
    }
  }
}

codesurface/
├── src/codesurface/
│   ├── server.py           # MCP server with 5 tools
│   ├── db.py               # SQLite + FTS5 database layer
│   ├── filters.py          # PathFilter (default exclusions, .codesurfaceignore, --exclude)
│   └── parsers/
│       ├── base.py         # BaseParser ABC
│       ├── cpp.py          # C++ header parser
│       ├── csharp.py       # C# parser
│       ├── go.py           # Go parser
│       ├── java.py         # Java parser
│       ├── python_parser.py # Python parser
│       └── typescript.py   # TypeScript/JavaScript parser
├── pyproject.toml
└── README.md

"No codebase indexed"

• Ensure --project points to a directory containing supported source files (.cs, .h, .hpp, .go, .java, .py, .ts, .tsx, .js, .jsx)
• The server indexes at startup. Check stderr for [codesurface] scanning N files... and [codesurface] done: lines

Server won't start

• Check Python version: python --version (needs 3.10+)
• Check mcp[cli] is installed: pip install mcp[cli]

Stale results after editing source files

• The index auto-refreshes on query misses. If you add a new class and query it, the server reindexes and retries automatically
• You can also call reindex() manually to force an incremental update

---

Contact

fuatcankoseoglu@gmail.com

License

MIT