mcp-gtags-server

Stop letting your AI agent grep. Give it an index.

mcp-name: io.github.harshithsunku/mcp-gtags-server

# If you don't have uv yet (one-time, no sudo):
curl -LsSf https://astral.sh/uv/install.sh | sh

{ "mcpServers": { "gtags": { "command": "uvx", "args": ["mcp-gtags-server"] } } }

One user-level config entry, no sudo, no pre-installed anything — the whole toolchain installs itself into user space on first use, and every repo you open is served automatically.

Every AI coding agent — Claude Code, Cursor, Codex, you name it — answers "where is this function defined?" the same way: grep the entire tree. On a million-line C/C++ codebase that's a full scan per question, and the output is a firehose: every comment, string literal, and unrelated match, dumped straight into the model's context window.

mcp-gtags-server replaces those scans with indexed lookups powered by GNU Global (gtags) — the same tags engine kernel and systems developers have trusted for decades — exposed to agents over the Model Context Protocol. Built for the codebases LSP-based tools can't handle: kernel-scale C/C++, trees that don't currently compile, machines you can't sudo on.

• ~100× faster per query — milliseconds instead of seconds, at any codebase size
• Radically less noise — the definition, not 7,873 lines of matches
• Speaks kernel — #ifdef guard stacks with .config filtering, macro-generated symbols (sys_read → its SYSCALL_DEFINE3 site) that no other tagging tool resolves, and definitions the parser misses recovered from their EXPORT_SYMBOL site via ctags
• Zero index management — first query builds the index, every query auto-refreshes it
• Correctness measured in CI — a 65-case golden eval covering all 11 tools against a pinned kernel: 100% recall, 100% precision@1 (docs/capability.md)
• Works everywhere MCP does — Claude Code, Claude Desktop, Cursor, any MCP client

The numbers (real Linux kernel, not a toy)

Measured on a full Linux kernel checkout — 65,163 C/C++ files, 37.1 million lines — warm page cache:

One-time index build: 66 s for the whole kernel. Incremental refresh after edits: well under a second. Reproduce it yourself with scripts/benchmark.sh:

./scripts/benchmark.sh /path/to/linux tcp_v4_rcv kmalloc ext4_readdir

The speed is nice. The real win is precision: an agent that gets 5 exact lines instead of 7,873 noisy ones keeps its context window for actual reasoning.

And the answers are measured, not assumed: CI runs a 65-case golden eval against pinned kernel v6.16 on every push — currently 100% recall, 100% precision@1 across definitions, macro resolution, export recovery, references, callers, callees, definition bodies, #ifdef guards, and reachability, covering all 11 tools. The full methodology, numbers, and honest limitations live in docs/capability.md.

Quick start (60 seconds)

Option A — one config entry (recommended)

Add the server once, at user level, and you're done — no installer, no pre-installed gtags, no per-repo setup. The only prerequisite is uv (or use Option B, which installs it for you).

# Claude Code (once per device, all repos):
claude mcp add --scope user gtags -- uvx mcp-gtags-server

// Cursor (~/.cursor/mcp.json), or any MCP client's global settings:
{
  "mcpServers": {
    "gtags": { "command": "uvx", "args": ["mcp-gtags-server"] }
  }
}

Or click the Cursor one-click install badge at the top of this page.

On the very first tool call the server bootstraps everything else by itself: GNU Global (prebuilt user-space binaries — no compiler, no sudo), universal-ctags, and Pygments, all into ~/.gtags-mcp. While that one-time install runs (a few seconds on most platforms), tool calls return a "toolchain is being installed — retry shortly" status instead of failing. Opt out with --no-auto-setup or GTAGS_MCP_AUTO_SETUP=0.

The prebuilt Linux binaries run on any distro with glibc ≥ 2.28 (RHEL/Rocky 8+, Ubuntu 18.10+, Debian 10+). On older hosts — or if a downloaded binary fails its post-install execution check — setup automatically compiles GNU Global from source instead, which only needs make and a C compiler.

Option B — one-line installer (shared background server)

One command. No sudo. Works everywhere — restricted corporate machines, containers, build servers:

curl -fsSL https://raw.githubusercontent.com/harshithsunku/mcp-gtags-server/main/scripts/install.sh | bash

Everything lands in your home directory — the server (via uv), GNU Global, universal-ctags, and Pygments (in ~/.gtags-mcp). When it finishes, a shared background HTTP server is running that every client and IDE window on the machine can point at (http://127.0.0.1:8383/mcp), and the exact client configuration is printed to your console.

Re-run the same command any time:

• Up to date? → "Already installed and up to date — nothing to install", and the config is printed again.
• New release on GitHub/PyPI? → the package updates, an outdated gtags toolchain is wiped and reinstalled automatically, and the background server restarts on the new version.

One config, many repos

However you install it, one user-level entry serves every repo you open — 20 repos need zero extra installs and zero extra config. Each tool call resolves its project root down this ladder:

1. project_root argument on the tool call (agents pass this to target any tree)
2. --root flag / GTAGS_MCP_ROOT env var
3. root in a config file (note: pinning a root here defeats multi-repo)
4. The client's workspace roots (MCP roots protocol) — IDEs that advertise their open folders get the right repo automatically, even on the shared HTTP server; with several folders open, agents are asked to pass project_root
5. Walk up from the server's working directory to the nearest .git/GTAGS — this is why stdio servers spawned by Claude Code/Cursor inside a repo just work

That's it. No indexing step, no configuration. Ask your agent "who calls tcp_v4_rcv?" — the first query in any repo builds that repo's index automatically, and every query after that is answered in milliseconds. Run mcp-gtags-server doctor any time to see what the server detects, or mcp-gtags-server config to re-print the client configuration.

# 1. GNU Global — EITHER user-space (no sudo):
mcp-gtags-server setup
#    OR a system package:
sudo apt install global      # Debian/Ubuntu
sudo dnf install global      # Fedora
brew install global          # macOS

# 2. The server:
uv tool install mcp-gtags-server        # or: pip install mcp-gtags-server

{
  "mcpServers": {
    "gtags": {
      "command": "uvx",
      "args": ["mcp-gtags-server", "--root", "/absolute/path/to/your/project"]
    }
  }
}

# .gtags-mcp.toml
label = "native-pygments"     # force a GTAGSLABEL parser label
bin_dir = "/opt/tools/bin"    # extra directory searched for gtags/global/ctags
skip_globs = ["*.gen.c"]      # never index paths/basenames matching these globs
respect_gitignore = true      # default: index only what `git ls-files` reports
enrich = true                 # default: ctags kind/signature/scope on results
guards = true                 # default: #ifdef guard stacks on results
macro_resolve = true          # default: resolve macro-generated symbols (sys_*, ...)
# root = "/abs/path"          # default project root (user config)

The tools

Symbol-level tools — the noise killers

Give the agent the symbol, not the file.

Core lookups

Every query tool supports limit/offset pagination, long-line truncation, and (where it makes sense) case_insensitive — output is engineered to never flood a context window.

Structured output (JSON by default)

Since v0.8.0 every tool returns a machine-readable JSON envelope by default (pass format="text" for the previous human-readable rendering — a breaking change if you parsed the old text):

{
  "tool": "find_definition",
  "root": "/abs/project/root",
  "results": [
    {"symbol": "kmap", "path": "include/linux/highmem-internal.h", "line": 40,
     "col": 22, "kind": "function", "typeref": "void *", "scope": null,
     "signature": "(struct page * page)", "guard": ["CONFIG_HIGHMEM"],
     "snippet": "static inline void *kmap(struct page *page)"}
  ],
  "total": 2, "offset": 0, "truncated": false,
  "next_tools": ["get_symbol_body", "find_callers", "symbol_info"],
  "warning": null
}

• Symbol locations always use the stable record schema {symbol, path, line, col, kind, typeref, scope, signature, guard, snippet} with repo-relative paths. Keys are only ever added, never renamed or removed — parsers never need to change shape.
• kind / typeref / scope / signature say what a symbol is (since v0.8.1): function vs. macro vs. struct vs. typedef vs. enum constant, its return/target type, its enclosing scope (enum:color, struct:item), and its parameter list — extracted per file by universal-ctags with no build and no compile database, cached, and filled on definition-shaped results (find_definition, symbol_info, list_file_symbols). When universal-ctags isn't available the fields are simply null; disable explicitly with --no-enrich, GTAGS_MCP_ENRICH=0, or enrich = false in .gtags-mcp.toml.
• guard says when a symbol exists (since v0.9.0): the enclosing #if/#ifdef stack, outermost first ([] = unconditional, null = scanning disabled or file unreadable). See the next section — this is the headline feature.
• resolved_via says how a symbol was found when it took macro-family resolution rather than a literal index match ("macro:SYSCALL_DEFINE", "fuzzy:vfs_read") — see Macro-generated symbols.
• next_tools tells the agent the highest-value follow-up call for what was (or wasn't) found.
• total/offset/truncated replace the text continuation footer; errors keep the envelope with an error field.
• Composite tools return tool-shaped results (e.g. find_callees {in_tree, external}, symbol_info an overview object, reachability a hop chain) inside the same envelope.

#ifdef-aware: know which definition your config actually compiles

Kernel and firmware code defines the same symbol multiple times and lets the build configuration pick one. Every other no-build tool returns a flat, unexplained list — an agent happily reads the no-op stub of kmap and reasons its way to a wrong answer. This server reads the preprocessor conditionals (pure scanning — still no build, no compile_commands.json):

Symbol: kmap
  2 definitions under 2 distinct guards:
  [CONFIG_HIGHMEM] defined at include/linux/highmem-internal.h:40 — function kmap(struct page * page) -> void *
  [!CONFIG_HIGHMEM] defined at include/linux/highmem-internal.h:170 — function kmap(struct page * page) -> void *

Pass active_config — a kernel .config path or a macro list like "CONFIG_SMP,BITS_PER_LONG=64,!CONFIG_DEBUG" — to find_definition, find_references, or symbol_info, and definitions whose guard stack is definitely false under it are dropped (the envelope reports the count as config_filtered). Filtering is deliberately conservative: a .config is a closed world for CONFIG_* macros (kbuild semantics, including =m → CONFIG_X_MODULE and IS_ENABLED/IS_BUILTIN/IS_MODULE), but anything unknown (__ASSEMBLY__, ARCH_HAS_*, arithmetic it can't decide) never drops a result.

The details are handled so the output stays clean: classic include guards (#ifndef FOO_H) are detected and suppressed, #elif chains compose into explicit conditions (!CONFIG_X86_64 && CONFIG_X86_32), comments on directives are ignored (they lie), and broken/partial files never fail a query. Disable with --no-guards, GTAGS_MCP_GUARDS=0, or guards = false in .gtags-mcp.toml.

Macro-generated symbols resolve too (sys_read → SYSCALL_DEFINE3)

The best-known gap of every tagging tool on the kernel: sys_read, trace_sched_switch, and css_set_lock have no literal definition anywhere — they are minted by token-pasting macros, and a plain lookup comes back empty (or worse, returns a same-named test helper). This server resolves them from the index alone, no preprocessor and no build:

find_definition("sys_read")        → fs/read_write.c:723  SYSCALL_DEFINE3(read, ...)   resolved_via: "macro:SYSCALL_DEFINE"
find_definition("__x64_sys_openat")→ fs/open.c:1385       SYSCALL_DEFINE4(openat, ...) (arch entry-point wrappers map back)
find_definition("trace_sched_switch") → include/trace/events/sched.h:220  TRACE_EVENT(sched_switch, ...)
find_definition("css_set_lock")    → kernel/cgroup/cgroup.c:82  DEFINE_SPINLOCK(css_set_lock);

Covered families: SYSCALL_DEFINE0..6 / COMPAT_SYSCALL_DEFINE* (incl. __x64_/__ia32_/__arm64_/__se_/__do_ wrapper spellings), tracepoints (TRACE_EVENT, DEFINE_EVENT, DECLARE_TRACE, ...), and the bare-name definers (DEFINE_SPINLOCK, DEFINE_MUTEX, DEFINE_PER_CPU*, DECLARE_BITMAP, module_param*, any DEFINE_/DECLARE_-shaped macro) — plus a last-resort fuzzy tier that tries underscore-variant spellings. Resolved results are flagged with resolved_via in the envelope and ranked ahead of same-named textual shadows; DEFINE_* sites rank above their DECLARE_* counterparts. symbol_info additionally reports the EXPORT_SYMBOL / EXPORT_SYMBOL_GPL variant a kernel symbol is exported with, in an exported field. Costs nothing when a symbol resolves normally (only family-shaped names like sys_*/trace_* get the extra indexed lookups); disable with --no-macro-resolve, GTAGS_MCP_MACRO_RESOLVE=0, or macro_resolve = false.

What gets indexed (junk stays out)

Indexing feeds gtags an explicit file list instead of letting it walk the tree:

• In a git repository the list comes from git ls-files — .gitignore is respected exactly, so build output, vendored blobs, and generated files never pollute the index. Disable with respect_gitignore = false in .gtags-mcp.toml.
• Outside git, the tree is walked minus well-known junk directories (.git, node_modules, build, dist, .venv, ...).
• skip_globs in .gtags-mcp.toml drops anything else you never want indexed.

Incremental refreshes recollect the list, so newly ignored files drop out of the index and new files appear — automatically.

The flow that saves your context window

1. symbol_info("kmalloc")                    → definitions + usage spread + next step (12 lines)
2. find_callers("ext4_mark_inode_dirty")     → deduped callers with counts (1 line/caller)
3. get_symbol_body("tcp_v4_rcv")             → read the ONE function that matters
4. find_callees("tcp_v4_rcv")                → what it depends on, with locations
5. reachability("ksys_read", "rw_verify_area") → the call chain, one line per hop
6. blast_radius("HEAD")                      → what my edit impacts, ranked by distance

A few hundred lines of context total — versus tens of thousands for the grep-and-read-files equivalent.

Multi-language projects (C + Python + more)

Real projects mix languages — a C core with Python tooling, JS frontends, Go services. The server handles this automatically:

• Native languages (C, C++, Java, PHP, Yacc, assembly) use GNU Global's fast built-in parser.
• Everything else (Python, Go, Rust, JavaScript, TypeScript, Ruby, ... ~150 languages) is indexed through Global's ctags + Pygments plugin parsers — same index, same tools, same queries.

The one-line installer (and mcp-gtags-server setup) enables this automatically — it installs universal-ctags and Pygments into user space, and the server switches to the native-pygments parser label on its own. Prefer system packages? Those work too:

sudo apt install exuberant-ctags python3-pygments   # Debian/Ubuntu
sudo dnf install ctags python3-pygments             # Fedora
brew install ctags && pip install pygments          # macOS

Now find_definition("py_util"), get_symbol_body (indentation-aware for Python), find_callees, find_callers — all work across every language in the tree, in one index.

Force a specific parser label with --label, GTAGS_MCP_LABEL, or label in .gtags-mcp.toml (e.g. default for native-only, pygments for plugin-everything).

Honest caveats: for plugin-parsed languages, definitions are as accurate as ctags, but references are token-based — every occurrence of the name counts, without C-grade semantic reference tracking or local-scope awareness. For C/C++ nothing changes: the native parser still does that part.

How it works

Architecture

One long-lived server process per IDE window (stdio) — or one shared HTTP server for the whole machine. Everything heavy lives on disk and is shared: the toolchain installs itself once per machine, the index builds itself once per repo.

flowchart LR
    subgraph clients["MCP clients"]
        CC["Claude Code / Cursor / VS Code"]
        CD["Claude Desktop (.mcpb)"]
    end

    subgraph proc["mcp-gtags-server — spawned once per window, lives for the session"]
        T["11 navigation tools"]
        RR["root resolution<br/>project_root → env → config → client roots → cwd"]
    end

    subgraph disk["your machine, user space (no sudo)"]
        TC["~/.gtags-mcp<br/>global · gtags · ctags · Pygments<br/>self-installs on first use"]
        IX["&lt;repo&gt;/.gtags-mcp/GTAGS<br/>one index per repo, auto-built"]
    end

    CC -- "stdio (uvx)" --> T
    CD -- "stdio" --> T
    T --> RR
    RR -- "runs global (ms, indexed)" --> TC
    TC -- "B-tree lookup" --> IX

A tool call, end to end

The binary is not re-executed per call — the server process stays alive; each call is one JSON-RPC message plus one millisecond-scale global subprocess:

sequenceDiagram
    participant A as Agent
    participant S as server (long-lived process)
    participant G as global (GTAGS index)

    A->>S: find_definition("tcp_v4_rcv")
    Note over S: first call on this machine?<br/>toolchain installs itself in the background,<br/>calls answer "installing — retry shortly"
    Note over S: first query in this repo?<br/>index auto-builds once (~66 s for the kernel)
    S->>G: global -dx tcp_v4_rcv
    G-->>S: net/ipv4/tcp_ipv4.c:2067 (milliseconds)
    S-->>A: narrow JSON answer — no grep firehose

• First query on a tree? The index is built automatically (the only operation that ever blocks — and only once).
• Where do the index files go? Into a single .gtags-mcp/ folder at the project root — never loose files next to your code. The folder ships its own .gitignore, so git status stays clean without touching yours. A pre-existing root-level GTAGS (from older versions, or your own gtags runs) keeps being used as-is. mcp-gtags-server doctor shows the location.
• Files changed? A debounced incremental refresh runs in the background: queries always answer instantly from the current index while gtags -i catches up behind the scenes. Measured on the kernel: queries return in 0.02s while the 25s freshness check runs invisibly. Staleness is bounded by the debounce window; call update_index for a synchronous, guaranteed-fresh barrier right after edits.
• Huge result? Pagination footers tell the agent exactly how to fetch the next page — or the tool itself suggests a narrower one (find_callers on a symbol used in 500+ files points to summarize_references).

FAQ

Why gtags instead of a language server (LSP)? LSP servers give richer semantics but need a working build configuration, per-editor setup, and serious warm-up time on large trees. gtags indexes 37M lines in about a minute with zero configuration, handles the kernel-scale codebases LSPs choke on, and its fuzzy parser doesn't care whether the code currently compiles. For C/C++ navigation questions — definition, references, callers — it's the pragmatic sweet spot. (Wrapping clangd for the compile-DB case was considered and deliberately rejected: users who have a working compile_commands.json already have clangd and its ecosystem — see ROADMAP.md.)

What languages? C, C++, Yacc, Java, PHP, and assembly natively — plus Python, Go, Rust, JS/TS, Ruby, and ~150 others via the ctags/Pygments plugin parsers (see Multi-language projects).

Does the agent have to manage the index? No. That's the point. Build-on-first-query, background refresh with adaptive debounce, zero blocking — queries never wait for index maintenance. The explicit update_index tool exists only as an escape hatch: a synchronous freshness barrier after edits, and a from-scratch rebuild with full=true.

Where does the index live? Can I delete it? In .gtags-mcp/ at the project root (self-gitignored). Delete it freely any time — the next query rebuilds it from scratch.

Will it fight my agent's built-in tools? The tool descriptions are written to steer the model: they say when to use indexed lookups instead of grep. In practice agents pick the faster, narrower tool naturally.

Development

git clone https://github.com/harshithsunku/mcp-gtags-server
cd mcp-gtags-server
uv run --extra dev pytest       # 229 tests; e2e tests auto-skip if GNU Global is absent
npx @modelcontextprotocol/inspector mcp-gtags-server    # poke at it interactively

Tests build a real C project in a temp dir and exercise auto-indexing, auto-refresh, caller mapping, body extraction, pagination, user-space binary discovery, and config layering end-to-end.

Correctness is also measured, not assumed: mcp-gtags-server eval --golden evals/golden.jsonl --root <kernel-tree> runs a 65-case golden set covering all 11 tools (definitions, macro resolution, references, callers, callees, bodies, guards, reachability, maintenance) against a real kernel and prints recall / precision@1 — CI does this weekly against a pinned tag. See docs/capability.md for the current numbers.

Release flow: bump version in pyproject.toml, tag vX.Y.Z, push — CI publishes to PyPI and users pick the update up on their next installer re-run. Prebuilt GNU Global binaries are rebuilt by tagging global-v<version> (or gh workflow run release-binaries.yml -f version=<version> to replace the assets of an existing release in place); Linux builds run inside manylinux_2_28 containers so they work on any glibc ≥ 2.28 host, enforced by a CI symbol-ceiling check.

Roadmap

See ROADMAP.md — structured JSON output landed in v0.8.0, ctags metadata enrichment (kind/signature/scope) in v0.8.1, #ifdef/config-guard awareness (the headline capability for kernel and firmware trees) in v0.9.0, then, all shipped in v1.0.0: macro-family symbol resolution (sys_read → its SYSCALL_DEFINE3 site), the agent workflow tools (reachability, blast_radius), automatic recovery from corrupted index databases, and the correctness eval harness — a 65-case golden set against pinned kernel v6.16 scoring 100% recall / 100% precision@1 in CI, with the measured writeup in docs/capability.md. Every technical milestone is done; what remains is distribution (MCP registry, directories, the writeup post).

Contributions welcome — open an issue or PR.

License

MIT © Harshith Sunku