Turjuman

Open-source, self-hosted translation management — managed by your AI agent.

Turjuman is a lightweight, open-source alternative to commercial translation-management SaaS. Instead of a heavy web dashboard, you manage projects, keys, locales and translations through an MCP server connected to Claude Code (or any MCP client/agent). A thin developer CLI handles the deterministic file work — pulling/pushing locale files in your repo and CI.

It runs on a serverless AWS stack (Lambda + DynamoDB) that costs ~nothing to host and fits comfortably in the free tier. No always-on servers, no Cognito.

Claude Code / agent ──Streamable HTTP + API key──┐
Developer CLI / CI ──REST + API key──────────────┤
                                                 ▼
                          Lambda Function URLs ──► DynamoDB (single table)

Why MCP-first?

Translation management is mostly conversation + judgement: "add a French locale", "translate the untranslated checkout strings", "what does cart.empty say in German?". An agent with the right tools does this naturally. The connected LLM does the translating itself — there's no separate machine-translation engine to pay for or configure.

The CLI exists for the things you should not hand to an LLM: deterministically exporting a JSON/YAML file, downloading translations, and syncing them in CI.

Related MCP server: openai-mcp

Features (v1)

• MCP server with 36 tools: projects, locales, keys (with descriptions/context), translations, bulk fill, review workflow, and full user/member/API-key administration (incl. key revocation).

• First-class RBAC. Global roles (OWNER/ADMIN/MEMBER) + per-project roles (MANAGER/EDITOR/DEVELOPER/VIEWER). High-privilege users manage everyone else's access.

• Glossary + translation memory, and webhooks (HMAC-signed, via DynamoDB Streams) for change events.

• API-key auth — simple bearer tokens, stored hashed. (OAuth/Cognito is a future option.)

• Developer CLI — login, init, pull, push, build with multi-target config and adapters for JSON (nested/flat), YAML, Flutter ARB, Java .properties, CSV, Android strings.xml, and iOS .strings/.stringsdict (ICU-canonical plurals converted to each format's native form).

• DynamoDB single-table design, on-demand billing. Multi-tenant-ready (orgId on every record).

Repository layout

Try it locally

The fastest way to see Turjuman work — everything runs against a local LocalStack (needs Docker + Node 24+), no cloud account needed. Turjuman uses pnpm (pinned via package.json's packageManager field); run corepack enable once to activate it.

pnpm install && pnpm run build
pnpm run localstack:up            # start the shared LocalStack on :4566
cp .env.example .env             # point the toolchain at LocalStack

# deploy into LocalStack with hot reload; prints the MCP/REST URLs + a fresh API key
pnpm run dev

pnpm run dev runs the real Lambda runtime in LocalStack (so DynamoDB Streams → webhooks fire too) and hot-reloads your edits. Point your MCP client at the printed MCP URL with the printed API key and start talking to it.

Self-host on AWS

Status: the CDK stack is conventional but has not yet been verified end-to-end against a live AWS account (see ROADMAP). Use the local path above to evaluate; treat the cloud deploy as beta and please report issues.

1. Deploy to your AWS account

Instantiate the TurjumanStack in a tiny CDK app and deploy it with the AWS CDK CLI. The construct vendors its Lambda bundles, so there's no SAM CLI and no repo clone (see Self-hosting for the full walkthrough and config):

npm install @turjuman/aws-cdk aws-cdk-lib constructs
cdk bootstrap   # once per account+region (standard CDK bootstrap)
cdk deploy      # prints McpUrl / ApiUrl / TableName as stack outputs

@turjuman/aws-cdk is standalone-installable, so the day-to-day turjuman developer CLI stays a lean, AWS-free install.

2. Your first owner + API key

Once the stack is up, create the first owner over HTTP and capture its key — printed once, and saved to ~/.turjuman/auth.json so you're logged in:

turjuman bootstrap --url <ApiUrl> --email you@example.com --name "Your Name"

It refuses to create a second owner once the deployment has users (returns 409), so re-running it is safe.

3. Connect Claude Code (MCP)

Add to your .mcp.json (see .mcp.json.example):

{
  "mcpServers": {
    "turjuman": {
      "type": "http",
      "url": "<McpUrl>",
      "headers": { "Authorization": "Bearer <your-api-key>" }
    }
  }
}

Keep your key out of version control. .mcp.json is often committed — prefer referencing the token via an environment variable / secret manager rather than pasting the literal key. If a key ever leaks, revoke it immediately with the revoke_api_key tool.

Then just ask: "Create a project called Web App with base locale en, add fr and es, and translate everything into French."

4. Sync files in your repo (CLI)

pnpm dlx turjuman login --url <ApiUrl> --key <your-api-key>
pnpm dlx turjuman init --project <proj_id> --format json-nested --path "locales/{locale}.json"
pnpm dlx turjuman pull     # write locale files from Turjuman
pnpm dlx turjuman push     # upload source keys / translations

Documentation

The full documentation lives in docs/ and is published as a documentation site with Mintlify — one source that serves human readers, external agents (Mintlify auto-generates llms.txt / llms-full.txt and an "open in Claude/ChatGPT" menu per page), and the agents working in this repo. Start here:

• Get started: Introduction · Quickstart (deploy to AWS) · Try it locally (no AWS)

• Concepts: Why MCP-first · Architecture · Lifecycle · Roles & permissions · How agents use Turjuman

• Guides: Translate with MCP · Code mode · Sync with the CLI · Quality checks · Webhooks · Connect Claude Code

• Reference: MCP tools · CLI commands · REST API · File formats · QA checks · Glossary

• Self-hosting: Overview · Deploy to AWS · Configuration · Security & API keys

• Contributing · Roadmap

Development

pnpm install
pnpm run build
pnpm test            # unit tests (hermetic)

# integration tests against an emulated AWS (LocalStack DynamoDB):
pnpm run localstack:up && pnpm run test:integration && pnpm run localstack:down

# full deployed end-to-end (SAM stack on LocalStack: Lambda Function URLs +
# the real DynamoDB Streams -> webhook flow). Just needs Docker:
pnpm run test:e2e

Formatting & linting

Formatting and linting are handled by Biome (a single, fast Rust-based replacement for Prettier + ESLint), configured in biome.json:

pnpm run check         # format + lint, report only
pnpm run check:write   # apply formatting + safe lint fixes
pnpm run format        # format only
pnpm run lint          # lint only

You rarely run these by hand. A pre-commit hook auto-formats staged files, CI runs biome ci ., and — since this repo is developed primarily with Claude Code — a PostToolUse hook (.claude/hooks/biome-format.sh) auto-formats each file as the agent edits it. The generated docs/api-reference/openapi.json snapshot is owned by gen:openapi and excluded from Biome.

See Contributing for the full testing guide (unit, LocalStack integration, and deployed end-to-end), which also runs in CI.

License

MIT