Turjuman
Open-source, self-hosted translation management system with an MCP server that enables AI agents to manage projects, locales, keys, and translations via natural language, and includes a CLI for deterministic file syncing.
README
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.
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,buildwith multi-target config and adapters for JSON (nested/flat), YAML, Flutter ARB, Java.properties, CSV, Androidstrings.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 (
orgIdon every record).
Repository layout
| Package | What it is |
|---|---|
packages/core |
Domain model, DynamoDB repository, RBAC, services, format adapters (the shared brain) |
packages/mcp-server |
Stateless Streamable HTTP MCP server → Lambda |
packages/api |
REST API for the CLI/CI → Lambda, plus the webhook dispatcher |
packages/cli |
The turjuman developer CLI (incl. the AWS CDK deployer in src/deploy/) |
Try it locally (no AWS)
The fastest way to see Turjuman work — everything runs against DynamoDB Local, no cloud account needed:
npm install && npm run build
docker run -d --rm -p 8000:8000 amazon/dynamodb-local
# create the table + first owner, and print your API key
node scripts/dev-setup.mjs you@example.com "You"
# start the servers (each in its own shell), pointed at DynamoDB Local
export AWS_ENDPOINT_URL_DYNAMODB=http://localhost:8000 TURJUMAN_TABLE=Turjuman
export AWS_ACCESS_KEY_ID=local AWS_SECRET_ACCESS_KEY=local AWS_REGION=us-east-1
node packages/mcp-server/dist/local.js # MCP on http://localhost:3000
node packages/api/dist/local.js # REST on http://localhost:4000
Point your MCP client at http://localhost:3000/ with the printed API key (see below) and start
talking to it.
Self-host on AWS
Status: the SAM 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
One command self-bootstraps the CDK environment, deploys the CloudFormation stack from pre-bundled Lambda assets, and creates your first owner — no SAM CLI, no clone needed (see Self-hosting for options):
npx @turjuman/aws-deploy deploy # interactive; prints McpUrl/ApiUrl and your API key
Self-hosting ships as a separate @turjuman/aws-deploy package (binary turjuman-aws-deploy) so the
day-to-day turjuman developer CLI stays a lean, AWS-free install.
2. Your first owner + API key
turjuman-aws-deploy deploy creates the first owner at the end and prints the API key once (also
saved to ~/.turjuman/auth.json, so you're logged in). It refuses to create a second owner once the
org has users, so re-running deploy to update the stack 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.jsonis 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 therevoke_api_keytool.
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)
npx turjuman login --url <ApiUrl> --key <your-api-key>
npx turjuman init --project <proj_id> --format json-nested --path "locales/{locale}.json"
npx turjuman pull # write locale files from Turjuman
npx 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:
- Introduction · Quickstart (run locally, no AWS) · Self-hosting (deploy on AWS)
- Architecture · Roles & permissions · Lifecycle
- Translate with MCP · Sync with the CLI · Quality checks · Webhooks
- Reference: MCP tools · CLI commands · File formats · QA checks
- Contributing · Roadmap
Development
npm install
npm run build
npm test # unit tests (hermetic)
# integration tests against an emulated AWS (LocalStack DynamoDB):
npm run e2e:up && npm run test:integration && npm run e2e:down
# full deployed end-to-end (SAM stack on LocalStack: Lambda Function URLs +
# the real DynamoDB Streams -> webhook flow). Just needs Docker:
npm run test:e2e
See Contributing for the full testing guide (unit, LocalStack integration, and deployed end-to-end), which also runs in CI.
License
MIT
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。