symbolhop

Go-to-definition and find-references for coding agents. Codex, Claude Code and Gemini CLI ask your language servers instead of grepping, and get exact answers in one short line each.

Drawn demo: asked where count_entries is used, grep returns 77 lines, 67 of them comments, log messages and a similar function; lsp_references returns 10 lines, every one a use.
Set it up Star on GitHub

free · open source · mcp server and cli · read-only · no dependencies

Five questions, answered exactly

  • lsp_references: every use of a symbol, without comments, strings or look-alikes.
  • lsp_definition: where it's defined.
  • lsp_hover: its type, signature and documentation.
  • lsp_diagnostics: errors and warnings in a file right after an edit.
  • lsp_symbols: a file's outline.

Set up in one line

  • Codex ~/.codex/config.toml: [mcp_servers.symbolhop] with command = "npx", args = ["-y", "github:stanczykwyatt/symbolhop#v0.1.0", "serve"]
  • Claude Code: claude mcp add symbolhop -- npx -y github:stanczykwyatt/symbolhop#v0.1.0 serve
  • Gemini CLI and Cursor: the same command under mcpServers.

Questions

Why not let the agent grep?

Grep returns comments, strings, log messages and similarly named symbols along with the real uses. In a measured example, it returned 77 lines for a function's uses, 67 of them not uses; lsp_references returned the 10 real ones.

Is it safe to give an agent?

It only asks read-only questions, only about files inside the project, never writes anything, has no network access and no dependencies. SECURITY.md has the details.

Which languages?

Any with a language server on your PATH that it knows: C, C++ and Objective-C with clangd, TypeScript and JavaScript, Python, Go, Rust, Swift, Java, Ruby, Lua and Zig.

Does Codex have this built in?

Not yet; openai/codex#8745 asks for it. symbolhop works with Codex today through MCP.