assemblix-mcp

MCP server for Assemblix. Author, run & inspect your project's workflows from any MCP client using a single project API key (sk_...). projectId is never needed — the key defines the project.

Get your sk_ key from the project's API keys page in the Assemblix UI.

Install

Hosted (easiest — nothing to install)

claude mcp add --transport http assemblix https://mcp.assmblx.com \
  --header "Authorization: Bearer sk_your_key"

Local via uvx (privacy / self-host)

claude mcp add assemblix \
  --env ASSEMBLIX_API_KEY=sk_your_key \
  --env ASSEMBLIX_API_URL=https://app.assmblx.com \
  -- uvx assemblix-mcp

Claude Desktop / Cursor (JSON config)

{
  "mcpServers": {
    "assemblix": {
      "command": "uvx",
      "args": ["assemblix-mcp"],
      "env": {
        "ASSEMBLIX_API_KEY": "sk_your_key",
        "ASSEMBLIX_API_URL": "https://app.assmblx.com"
      }
    }
  }
}

For the hosted server, use your client's "add custom connector → URL + header" flow.

Configuration

Tools

Workflow authoring: list_node_types, list_workflows, get_workflow, create_workflow, update_workflow, publish_workflow Run & inspect: run_workflow, run_workflow_and_wait, get_execution, list_executions, get_execution_detail, list_in_flight Project state: list_project_state_variables, set_project_state_variables, upsert_project_state_variable, delete_project_state_variable Voice agents: list_conversation_voices, list_cascade_options, list_avatars, list_voice_agents, get_voice_agent, create_voice_agent, update_voice_agent, delete_voice_agent, list_voice_calls, get_voice_call Credentials: list_credentials (id, name and type only — never the secret)

Resources: assemblix://examples/minimal, assemblix://examples/branching, assemblix://examples/jev-routing (example workflow JSON — the last one validates and routes messages with a jev node), assemblix://guides/execution (how to call a workflow from your product — sync/async/streaming/sessions/voice, with curl/JS/Python examples), assemblix://guides/voice-agents (what a voice agent is, the call WebSocket protocol, analysis hooks, reading calls back, avatar calls over LiveKit, cascade mode). Prompts: author_workflow, integrate_workflow, integrate_voice_agent.

Workflow lifecycle: create_workflow → update_workflow(nodes, edges) → publish_workflow → run_workflow. Runs always execute the published snapshot. AGENT nodes need a provider credential configured in the Assemblix UI; jev nodes need a TypeSafe credential id, which list_credentials(type="typesafe_token") returns.

A voice agent is a different thing: no graph, no publish step — a prompt plus a voice, answering a live call. Workflows attach to it only as background analysis. The voice is either one speech-to-speech model (realtime, built with list_conversation_voices → create_voice_agent) or a cascade of streaming STT → a text LLM → streaming TTS (list_cascade_options → create_voice_agent(mode="cascade")), which is cheaper and stronger in Russian. Iterate on real calls with list_voice_calls / get_voice_call.

All tools operate on the single project your API key is scoped to; projectId is never a parameter.