mcp-locator

Point the server at one or more JSON files with --locale-<CODE>=<PATH> and it dynamically creates a set of tools for every locale. The AI sees set_en, get_fr, tree_de, search_uk, etc. and knows immediately which language each tool targets.

💸 Why — token savings on large catalogs

The usual way an agent edits translations is to read the entire locale file into context, edit, and write it back. That cost scales with the size of your catalog and repeats on every turn.

mcp-locator instead exposes targeted tools — the agent fetches only the keys it needs (get, search, diff) and writes only what changed (set_batch). The full catalog never enters the context window.

Rough estimate (≈ 18–22 tokens per "key.path": "value" entry, measured on typical nested JSON):

Numbers are estimates for illustration — actual usage depends on key-path length, value length, and how many keys a task touches. The point holds: cost grows with what you change, not with how big the catalog is.

📦 Install

Global (use as a CLI):

npm install -g mcp-locator
mcp-locator --locale-en=./locales/en.json --locale-fr=./locales/fr.json

Without installing (npx):

npx mcp-locator --locale-en=./locales/en.json --locale-fr=./locales/fr.json

From source:

git clone https://github.com/Tankonyako/mcp-locator.git
cd mcp-locator
npm install
node bin/mcp-locator.js --locale-en=./locales/en.json

🤖 Agent / MCP client setup

Most agents just need a command + args pointing at mcp-locator. Use absolute paths in global configs so the server finds files regardless of the working directory.

Claude Code

One-liner:

claude mcp add locator -- npx mcp-locator \
  --locale-en=./locales/en.json --locale-uk=./locales/uk.json

Or commit a project-scoped .mcp.json to the repo root:

{
  "mcpServers": {
    "locator": {
      "command": "npx",
      "args": [
        "mcp-locator",
        "--locale-en=./locales/en.json",
        "--locale-uk=./locales/uk.json",
        "--mode=pretty"
      ]
    }
  }
}

Gemini CLI

Add to ~/.gemini/settings.json (or a project .gemini/settings.json):

{
  "mcpServers": {
    "locator": {
      "command": "npx",
      "args": [
        "mcp-locator",
        "--locale-en=./locales/en.json",
        "--locale-uk=./locales/uk.json"
      ]
    }
  }
}

Codex CLI

Add to ~/.codex/config.toml:

[mcp_servers.locator]
command = "npx"
args = [
  "mcp-locator",
  "--locale-en=./locales/en.json",
  "--locale-uk=./locales/uk.json",
]

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "locator": {
      "command": "npx",
      "args": [
        "mcp-locator",
        "--locale-en=/absolute/path/to/locales/en.json",
        "--locale-fr=/absolute/path/to/locales/fr.json"
      ]
    }
  }
}

🧭 Practical workflow: English as the source of truth

A common setup is "always write in English while you build, then translate the missing locales later in one pass." Drop an instruction block like this into your agent's AGENTS.md / CLAUDE.md / system prompt:

You manage translations through the `locator` MCP server. Rules:

1. ALWAYS write user-facing strings in English (`en`) — and ONLY `en`.
   English is the single source of truth. Use set_en / set_batch_en.
2. Do NOT translate into other locales while building a feature. Leave the
   other locales alone — they get filled in later, on demand (see below).
3. Never hand-edit the locale JSON files. Always go through the tools.
4. Keys are namespaced by domain: `OrderTopUp.minDepositTitle`, not full
   sentences. Search before inventing — reuse beats duplicate (search_en).

Then, once the English copy is finished, you ask for translations as a separate step. A natural-language command like this scans for everything missing and fills it in across every locale and phrase:

"Sync translations: run diff to find keys missing from each locale, read the English source for each with get_en, translate them, and write them back with set_batch_<code>."

Under the hood the agent runs:

1. diff → for each non-base locale, the list of keys present in en but missing.

2. get_batch_en → the English source strings for those keys (one call).

3. translate them, then set_batch_<code> → write all translations per locale in one call.

Because only the missing keys move through the context, syncing a 2,000-key catalog costs a few thousand tokens, not a few hundred thousand.

🗂️ Works with any locale JSON — nested or flat

mcp-locator reads any JSON object, whether your keys are nested objects or a flat map. The on-disk shape is controlled by --mode; tool inputs always use the same dotted key paths (user.profile.name).

Nested keys (--mode=pretty, the default):

{
  "app": { "title": "My App" },
  "user": {
    "email": "Email address",
    "profile": { "name": "Name", "age": "Age" }
  }
}

Flat keys (--mode=inline):

{
  "app.title": "My App",
  "user.email": "Email address",
  "user.profile.name": "Name",
  "user.profile.age": "Age"
}

Both files expose the exact same key paths to the AI. Pick whichever your project already uses — point the server at it and go.

✨ Ideal for next-intl & one-file-per-locale projects

mcp-locator fits perfectly with next-intl, i18next, react-i18next, vue-i18n, and any setup where each locale lives in a single JSON file (messages/en.json, messages/uk.json, …). That's the canonical layout these libraries use, and it's exactly what --locale-<CODE>=<PATH> maps onto:

mcp-locator \
  --locale-en=./messages/en.json \
  --locale-uk=./messages/uk.json

🛣️ Roadmap: today each locale is one JSON file (one file = one locale's full catalog). Pointing a locale at a folder of split message files (e.g. messages/en/*.json namespaced per file) is planned — for now, consolidate to a single file per locale.

🚩 CLI flags

# Two locales, default pretty mode with dot splitter
mcp-locator --locale-en=./en.json --locale-fr=./fr.json

# Three locales, inline mode, double-underscore splitter
mcp-locator --locale-en=./en.json --locale-fr=./fr.json --locale-de=./de.json \
  --mode=inline --splitter=__

🛠️ Tools (per locale)

For each locale <CODE>, the server registers these tools:

set_<CODE>

Set (create or overwrite) a translation value. Writes to disk atomically.

{ "ok": true, "locale": "en", "key": "user.profile.name", "value": "Alice" }

get_<CODE>

Get a translation by key. Throws if the key does not exist.

{ "locale": "en", "key": "user.profile.name", "value": "Alice" }

has_<CODE>

Check whether a key exists.

{ "locale": "en", "key": "user.profile.name", "exists": true }

list_<CODE>

Return every key in the locale, sorted alphabetically.

{ "locale": "en", "count": 3, "keys": ["app.title", "user.email", "user.profile.name"] }

tree_<CODE>

Render the locale keys as an ASCII tree, reconstructed from the key paths.

depth: true — the whole tree:

EN locale — 4 key(s)
├─ app
│  └─ title
└─ user
   ├─ email
   └─ profile
      ├─ age
      └─ name

depth: false — structure only (string leaves hidden — great for a quick map of a large catalog):

EN locale — 4 key(s) (structure only)
├─ app
└─ user
   └─ profile

search_<CODE>

Substring-search key paths. Returns all matching key/value pairs.

{
  "locale": "en",
  "query": "profile",
  "count": 1,
  "matches": [{ "key": "user.profile.name", "value": "Alice" }]
}

Batch tools (per locale)

has_batch_<CODE>, get_batch_<CODE>, set_batch_<CODE> — the same operations across many keys in a single call. set_batch_<CODE> writes all entries and saves the file once at the end.

Cross-locale tools

• get_all — get one key's value across every loaded locale at once.

• set_all — set one key in many locales in a single call (pass { "en": "Hello", "fr": "Bonjour" }).

• diff — compare every locale against the base (first-loaded) locale; reports missing and extra keys per locale. This is what powers the "sync translations" workflow above.

🧪 Testing with MCP Inspector

npm install
npx @modelcontextprotocol/inspector node bin/mcp-locator.js -- \
  --locale-en=/tmp/en.json --locale-fr=/tmp/fr.json --mode=pretty

Open the Inspector URL shown in the terminal to call tools interactively.

🔧 Troubleshooting

• Server logs appear in tool output — logs go to stderr, not stdout. stdout is the JSON-RPC channel. Check your client's stderr pane.

• File not found on startup — the server starts with an empty locale and creates the file on the first set_* call.

• Atomic writes — all saves write to a .tmp file first, then rename, so a crash mid-write never corrupts your locale file.

• Non-string values in existing files — numbers, booleans, and arrays are coerced to strings with a warning logged to stderr.

🐛 Issues & contributing

Hit a bug, or want a feature? 🙌 Open an issue at https://github.com/Tankonyako/mcp-locator/issues — please include your --mode, your --splitter, and a small snippet of the locale file if it's relevant. 💬

Pull requests are very welcome! 🚀 Fork, branch, and open a PR against main.