symbolpeek-mcp

Semantic code intelligence for AI coding agents

Ask for the symbol you need—not the entire file.

SymbolPeek is an MCP server that gives an AI coding agent a symbol-level view of your codebase. Instead of reading a whole file to answer "what does this function do" or "who calls it", the agent asks for one declaration and gets exactly that.

TypeScript and JavaScript are analyzed with the official TypeScript Compiler API. Rust, Python, Java, Go, JSON, and Markdown use embedded Tree-sitter for syntax-level operations.

Quick start

Requirements

• An MCP client. Any stdio MCP client works; Codex and Claude Code are documented below.

• Node.js 20 or newer — required for .ts, .tsx, .js, and .jsx analysis. Rust, Python, Java, Go, JSON, and Markdown work without it.

No clone, Rust toolchain, or manual build is required.

macOS or Linux

curl -fsSL https://raw.githubusercontent.com/pioner92/symbolpeek-mcp/main/scripts/install.sh | sh

Windows PowerShell

irm https://raw.githubusercontent.com/pioner92/symbolpeek-mcp/main/scripts/install.ps1 | iex

The installer verifies the release checksum, installs the package, and then prints the exact command to connect your client. It touches nothing outside its own install directory — no client configuration is modified for you.

That is the complete list. Connecting the MCP server and installing the optional agent guidance are separate, explicit steps. See Uninstall to remove everything.

To inspect the installer before running it:

curl -fsSL https://raw.githubusercontent.com/pioner92/symbolpeek-mcp/main/scripts/install.sh -o install-symbolpeek.sh
less install-symbolpeek.sh
sh install-symbolpeek.sh

Connect your client

The installer prints these commands with the correct absolute path already filled in. Use the absolute path form — it works regardless of how your client inherits PATH:

# Codex (macOS/Linux)
codex mcp add symbolpeek -- "$HOME/.local/share/symbolpeek/symbolpeek"

# Claude Code (macOS/Linux)
claude mcp add --transport stdio --scope user symbolpeek -- "$HOME/.local/share/symbolpeek/symbolpeek"

# Codex (Windows)
codex mcp add symbolpeek -- "$env:LOCALAPPDATA\SymbolPeek\symbolpeek.exe"

# Claude Code (Windows)
claude mcp add --transport stdio --scope user symbolpeek -- "$env:LOCALAPPDATA\SymbolPeek\symbolpeek.exe"

Restart the client, then verify: codex mcp list, or /mcp inside Claude Code. Use --scope project with Claude Code to enable the server for one project only.

Optional: agent guidance

The tools work as soon as the server is connected. SymbolPeek also ships a short skill that tells the agent to reach for those tools before opening whole files — without it the model calls them less often.

Install it only for the clients you actually use:

symbolpeek install-skills codex     # writes ~/.codex/skills/symbolpeek
symbolpeek install-skills claude    # writes ~/.claude/skills/symbolpeek
symbolpeek install-skills all       # both

CODEX_HOME and CLAUDE_CONFIG_DIR override those locations. Restart the client afterwards so it discovers the skill.

Try it with a prompt like:

Use the symbolpeek MCP server. List the symbols in the absolute path
/project/src/dashboard.tsx, then read_symbol_context for Dashboard.
After that, find_references for useAuth and go_to_definition for one usage.

Configuration templates are checked in at config/codex-mcp.toml.example and config/claude-mcp.json.example.

Related MCP server: @aiready/ast-mcp-server

What you get

A real request against this repository's own TypeScript worker (src/language/typescript/worker.js, 1,791 lines / 65 KB):

read_symbol(path: ".../src/language/typescript/worker.js",
            symbol: "createProject.collectImports")

{
  "symbol": "createProject.collectImports",
  "kind": "function",
  "file": ".../src/language/typescript/worker.js",
  "lines": { "start": 713, "end": 751 },
  "source": "function collectImports(fileName, collected, visited) {\n  ...\n}",
  "analysis": { "backend": "ts-compiler-api", "analysis_level": "syntax", "complete": true }
}

The agent receives 2.0 KB instead of 65 KB — the 39 lines it asked for rather than 1,791 lines it did not. Nested symbols are addressable by qualified name, so the agent never has to read the enclosing function to reach the inner one.

Language support

Every language-aware operation, exactly as reported by get_capabilities:

¹ Same-file only — conservative, no cross-file resolution. ² Explicit impl Type and impl Trait for Type blocks. Alias, re-export, and blanket-impl resolution requires rust-analyzer and is not supported.

Unsupported operations fail with an explicit error rather than returning empty results. get_capabilities returns this same matrix at runtime.

Tools

Navigation

Code intelligence

Type analysis

Refactoring

Discovery

Every request shape, option, and response format is documented in the MCP tool reference.

To reduce the MCP startup context, disable tools that your workflow does not need:

sym tools                              # interactive menu
sym tools list
sym tools disable get_type get_diagnostics
sym tools enable get_type
sym tools reset                        # enable everything

Changes take effect after restarting the MCP server. Disabled tools are omitted from tools/list, rejected if a client calls a cached schema directly, and removed from get_capabilities. Historical counters remain visible through sym usage.

Where it helps (and where it doesn't)

SymbolPeek does not replace reading files or text search. It replaces the most expensive pattern an agent hits on a real codebase: find a symbol, find everyone who touches it, and understand its type — across import aliases and re-exports.

Where it genuinely saves an agent work and tokens

• One symbol out of a large file. read_symbol / read_symbol_context return a single declaration instead of a whole file. On large files this is the bulk of the token saving.

• Impact analysis without a grep loop. find_references returns resolved locations, while find_calls with direction: "callers" also identifies the enclosing functions in one call. Both use the project's module resolution, including path aliases (@app/...) and barrel re-exports, which text search follows unreliably.

• Resolved types. get_type returns the instantiated signature (for example useAsync<FeedPermission, Error>), which text search cannot produce.

• Scoped discovery. search_symbols finds declarations by name and kind (for example "hooks matching conversation") without the noise of string matches in comments and unrelated code.

Where an agent should still use ordinary tools

• Plain text, comments, small config files, or unsupported files — grep is faster.

• Understanding the full control flow inside one function — just read it.

• Very large monorepos — building a TypeScript program has real latency, so a targeted grep can win for a single lookup. Rust, Python, Java, Go, and JSON requests never start Node and stay fast regardless of project size.

What makes the results trustworthy

• Parsing, source ranges, type information, and cross-file navigation come from the TypeScript compiler and its AST. Semantic labels such as hook and react_component are conventions applied on top of that syntax tree, based on naming and JSX usage.

• Cross-file results are only as complete as module resolution allows: with a valid tsconfig.json they use its configured source set; without one they cover the target file and recursively resolved static imports, exports, and require(...) calls. Compiler options come from the project tsconfig.json.

• Rust, Python, Java, Go, and JSON source ranges and nesting come from Tree-sitter. Every result carries compact analysis: { backend, analysis_level, complete } trust metadata, where complete: false means the parser recovered from a syntax error.

What gets indexed

TypeScript and JavaScript — function declarations, async functions, generators, and arrow functions; exported and nested functions; React components and hooks; classes and class methods; object methods; interfaces, type aliases, enums and qualified enum members, variables, and constants.

Enum members are addressed by qualified name, for example Screens.PUBLISH_ACKNOWLEDGEMENT. Symbol names are indexed; assigned string values remain literals and require text search.

Rust — functions, structs, unions, enums and variants, traits, impl blocks and methods, modules, constants, statics, type declarations, and macros. Impl methods use qualified names such as Client.send, and trait impl methods use <Client as Transport>.send.

JSON — object properties are indexed as RFC 6901 JSON Pointers, for example /checkout/errors/payment_failed. Array-valued properties remain single addressable branches instead of expanding every element, which keeps large locale and data files token-efficient. .jsonc and JSON5 are not supported.

Markdown — headings are the symbols, nested by level, and a symbol spans the whole section rather than the heading line. read_symbol with Quick start.Connect your client returns that section alone; on this README that is 839 bytes instead of 20 KB. Both # and underline (setext) headings are indexed, # inside a fenced code block is not, and repeated headings such as Options under several commands get @line:column selectors so each stays addressable. Prose, lists, and code blocks are not indexed separately — this finds sections, not full-text matches.

Using the tools

Absolute file paths are the canonical, most reliable input. Relative paths first use an explicit SYMBOLPEEK_WORKSPACE_ROOT, then filesystem roots supplied by a compatible MCP client; multi-root workspaces resolve only when exactly one root contains the requested path.

Supported files are parsed from their current contents on every request — there is no index to rebuild and no stale cache to invalidate.

Unsupported extensions return { "supported": false }. Missing files, parser failures, unknown symbols, and paths that are not usable source files (a directory, a bare module specifier) are returned as MCP invalid-parameter errors that name the mistake.

How agents discover the tools

Two mechanisms nudge a model toward targeted reads:

• Server instructions, always on. Every MCP initialization response includes concise instructions to inspect outlines or search first, then retrieve only the required symbol. This works with any MCP client that exposes server instructions to its LLM — nothing to install.

• The bundled skill, opt-in. symbolpeek skill is a stronger, always-loaded hint for Codex and Claude Code, with a trigger description covering code exploration and large JSON locale/configuration files. Install it with install-skills if you want it.

No MCP server can force a client model to call a tool, but these mechanisms make the intended workflow part of the model's default context. For another agent that consumes neither, copy the short workflow from the bundled skill into that client's global agent instructions.

Statistics

The CLI reports lifetime context-avoidance statistics:

symbolpeek stats
symbolpeek stats --reset

symbolpeek stat is accepted as a shorter alias for stats.

Tool-call frequency is persisted separately in the same statistics file:

symbolpeek usage
symbolpeek usage --reset

usage lists every current MCP tool, including zero-call tools, ordered by descending call count. Calls rejected for invalid parameters still count because they show attempted use; unknown tool names do not. Each --reset clears only the selected report, so stats --reset preserves usage history and vice versa. Concurrent Codex, Claude, and Copilot MCP processes merge their increments into the same file without overwriting one another. After either reset command, already-running servers add only post-reset deltas and cannot restore old totals.

All numbers compare SymbolPeek with a counterfactual full-source baseline:

• requests — successful semantic calls;

• files avoided — distinct source files represented by each result, summed across calls (one find_calls request can count several files);

• bytes / lines avoided — full contents of those files minus a compact serialization of the semantic result;

• estimated token savings — avoided bytes at a fixed ~4 bytes/token heuristic;

• average context reduction — size-weighted across all requests.

Treat these as directional context-reduction estimates, not exact model-token counts or billing data.

Configuration

Prebuilt binaries detect their bundled TypeScript runtime automatically. These variables are for advanced setups only:

For a global MCP installation, do not set SYMBOLPEEK_WORKSPACE_ROOT to a fixed project — use absolute paths, or let the client provide filesystem roots. Set SYMBOLPEEK_ALLOW_CWD_FALLBACK=false if relative paths must never resolve against the server's working directory.

Troubleshooting

The client shows no SymbolPeek tools. Confirm registration with codex mcp list or /mcp in Claude Code, and restart the client — MCP servers are discovered at startup. If registration itself failed, re-run the mcp add command using the absolute binary path.

TypeScript/JavaScript calls fail, other languages work. Node.js 20+ is missing or not visible to the server process. Check with node --version; if Node is installed but not on the client's PATH, set SYMBOLPEEK_NODE to the absolute Node executable.

symbolpeek: command not found after installing on macOS/Linux. ~/.local/bin is not on your PATH:

export PATH="$HOME/.local/bin:$PATH"   # add to your shell profile

The absolute path ~/.local/share/symbolpeek/symbolpeek always works.

Windows: the command is not recognized. The installer updates the user PATH; open a new terminal so the change is picked up. If irm ... | iex is blocked, run Set-ExecutionPolicy -Scope Process RemoteSigned first, or download the archive manually.

macOS blocks the binary. Release binaries are currently unsigned. If a browser-added quarantine flag blocks a manually downloaded package, clear it:

xattr -dr com.apple.quarantine <package-directory>

A relative path resolves to the wrong project. Use absolute paths, or set SYMBOLPEEK_WORKSPACE_ROOT for a deliberately project-scoped launch. SYMBOLPEEK_ALLOW_CWD_FALLBACK=false disables working directory fallback entirely.

An operation returns an unsupported-operation error. Check the language support matrix — semantic operations are TypeScript/JavaScript only.

Still stuck? Open an issue with your OS, client, SymbolPeek version (symbolpeek --version), and the failing call.

Uninstall

# macOS / Linux
rm -rf ~/.local/share/symbolpeek ~/.local/bin/symbolpeek ~/.local/bin/sym
rm -rf ~/.config/symbolpeek "$HOME/Library/Application Support/SymbolPeek"

# Windows
Remove-Item -Recurse -Force "$env:LOCALAPPDATA\SymbolPeek", "$env:APPDATA\SymbolPeek"

Then remove the server from your client: codex mcp remove symbolpeek or claude mcp remove symbolpeek. On Windows, also drop the SymbolPeek entry from your user PATH.

If you installed the optional agent guidance, remove it too:

rm -rf ~/.codex/skills/symbolpeek ~/.claude/skills/symbolpeek

Direct downloads

Prefer to install manually? Every package contains symbolpeek, the sym alias, and the locked TypeScript runtime.

Every package has a matching .sha256 asset. All versions and release notes are on the GitHub Releases page.

A manually extracted archive behaves exactly like an installer-placed one: run the binary directly, register it with your client, and optionally run install-skills.

SymbolPeek communicates over stdio when used as an MCP server. It normally does not print a terminal interface; an MCP client starts it and exchanges JSON-RPC messages through stdin/stdout.

Documentation

Roadmap

The current foundation is intentionally focused. Natural next capabilities include:

• deep type expansion beyond get_type hover signatures (fully resolved nested and generic types);

• JSX component trees and prop-flow analysis;

• project indexing and incremental parsing;

• additional language providers (Kotlin, Swift, C++).

License

MIT © Alex Shumihin