2kw.ai MCP Server
An MCP (Model Context Protocol) server for 2kw.ai — the EU-hosted AI integration platform. It gives AI coding assistants like Claude Code, Cursor, and Windsurf direct access to an OpenAI-compatible LLM gateway, schema-driven document extraction, audio transcription, custom agents with a knowledge base, and cost observability — 199 tools, straight from your editor.
Typical prompts once installed: "extract the line items from this invoice PDF", "create a schema for purchase orders and test it against these samples", "transcribe this call recording", "what did our extraction traffic cost this week?".
Installation
Claude Code
claude mcp add 2kw --env AI_2KW_API_KEY=sk_your_api_key -- npx -y @2kw/ai-mcp-server
Cursor
Add to .cursor/mcp.json in your project root:
{
"mcpServers": {
"2kw": {
"command": "npx",
"args": ["-y", "@2kw/ai-mcp-server"],
"env": {
"AI_2KW_API_KEY": "sk_your_api_key"
}
}
}
}
Windsurf
Add the same block to .windsurf/mcp.json. Any MCP-compatible client works with this configuration pattern — the server speaks stdio by default and streamable HTTP when MCP_TRANSPORT=http.
Get an API key from your 2kw.ai dashboard — paid plans start with a 7-day free trial.
Environment Variables
| Variable | Required | Default | Description |
|---|---|---|---|
AI_2KW_API_KEY | Yes | — | API key for the 2kw.ai platform (legacy alias: BACKBONE_API_KEY) |
AI_2KW_BASE_URL | No | 2kw.ai cloud API | Point at a self-hosted deployment (legacy alias: BACKBONE_BASE_URL) |
AI_2KW_MEMORY | No | off | true makes agent runs through 2kw_create_response and 2kw_decide_agent_approvals opt into agent memory (X-Backbone-Memory: enabled) |
MCP_TRANSPORT | No | stdio | Transport mode: stdio or http |
MCP_HTTP_PORT | No | 3100 | Port for HTTP transport |
Available Tools
| Category | What your assistant can do |
|---|---|
| Document conversion | Convert PDF, DOCX, XLSX, images, and 20+ formats to Markdown/text/HTML/JSON |
| Data extraction | Run schema-driven extractions (sync/async), estimate tokens, re-run, inspect results |
| AI Gateway | Chat completions and Responses (agent or direct model) through the OpenAI-compatible gateway, list available models |
| Transcription | Transcribe audio (FLAC, MP3, MP4, OGG, WAV, WebM) to text, JSON, SRT, VTT |
| Schemas | CRUD, versioning, labels, validation, and testing against sample documents |
| Prompts | Versioned prompt management, labels, compilation, testing |
| Datasets & experiments | Build datasets, run experiments, manage evaluators, record scores |
| Agents | CRUD, versioning, labels, runs with the full result envelope, approval decisions, tool-approval history, and tool-catalog sync history |
| Conversations | Create/inspect/delete conversations and replay their items for the responses API |
| Knowledge base | CRUD, document upload/attach/versioning, hybrid search, and citation resolution |
| Files | Upload, list, download, and delete files (agent input or knowledge documents) |
| Memory | List, read, write and delete your own agent memory files, forget everything; admins see member totals and erase a member |
| Observability | Analytics (usage, quality, providers, errors), traces, billing limits |
| API docs | Browse the platform API documentation by section |
Why 2kw.ai
- EU data residency — built-in models run GDPR-compliant and EU-hosted, operated by Manfred Kunze Development GmbH (Germany)
- One platform — gateway, extraction, transcription, and observability share one API, one dashboard, one bill
- Grounding scores — every extracted value links back to where it came from in the source document
Related
- 2kw.ai — platform overview and pricing
- Documentation — full API reference and guides
@2kw/ai— the 2kw.ai CLI (2kw,backbone,bb) for terminals and agentic workflows
License
Proprietary — see LICENSE for details.