Tendem MCP for agent builders. Embed humans into your pipelines

Tendem takes on the tasks too complex for AI alone — the ones you won’t trust an agent to finish end-to-end, where a human lifts the quality. At the core is an orchestrator: it analyzes a task, breaks it into steps, and routes each step to the right executor — an AI agent, a human expert, LLM-QA, or human-QA — then composes the outputs and decides what to do next.

How it works. A remote MCP server (portable across agents and platforms). Your agent delegates a task it can’t finish alone and gets back a finished, QA’d artifact — with a vetted human in the loop and in control of scope and approval. Results come back as markdown + files, straight into the flow. Nothing sensitive crosses MCP. Billed from your Tendem balance.

What you can delegate today: research, data enrichment, consulting, crowd-testing, and expert reviews/assessments/opinionated judgment (tech, design, copywriting). Recently expanded into STEM (life sciences, clinical, niche experts, finance) and phone calls.

Why it’s interesting for pipelines: anywhere your automation needs a human in the loop — judgment calls, verification, data the model can’t reach, quality review, edge cases — the agent calls a vetted Tendem expert through the connector instead of stalling or guessing, and a QA’d artifact lands back in the flow. No custom backend, no queue to build. It’s one MCP connector.

Getting started

Step 1 — Create your account. Sign in at https://agent.tendem.ai and go through the short onboarding.

Once you’re in the product, the full MCP connection instructions live at https://agent.tendem.ai/mcp — under the Agent builders tab.

Connect the MCP (for pipelines / agent builders)

Server-to-server access — no OAuth, no browser. This is the path for embedding Tendem into a pipeline or your own agent runtime.

Step 2 — Generate an API key — at https://agent.tendem.ai/mcp → Agent builders tab → Generate API key. The key is tied to your Tendem account.

Step 3 — Point your MCP client at the server URL and send the key in the Authorization header. Works with any client that speaks remote HTTP with custom headers. Typical config:

{
  "mcpServers": {
    "tendem": {
      "type": "http",
      "url": "https://mcp.tendem.ai/mcp?utm_hash=1fc8837791",
      "headers": {
        "Authorization": "ApiKey <your-api-key>"
      }
    }
  }
}

{
  "mcpServers": {
    "tendem": {
      "type": "http",
      "url": "https://mcp.tendem.ai/mcp?utm_hash=1fc8837791",
      "headers": {
        "Authorization": "ApiKey <your-api-key>"
      }
    }
  }
}

Python / LangChain: skip the plumbing

On LangChain you don’t have to wire the MCP loop yourself — langchain-tendem wraps Tendem into four tools governed by a spend cap:

pip install langchain-tendem
export TENDEM_API_KEY=...   # agent.tendem.ai/mcp → "Agent builders" tab

pip install langchain-tendem
export TENDEM_API_KEY=...   # agent.tendem.ai/mcp → "Agent builders" tab

from langchain.agents import create_agent
from langchain_tendem import tendem_tools

agent = create_agent("anthropic:claude-sonnet-4-5", tools=tendem_tools(max_price=25.0))

from langchain.agents import create_agent
from langchain_tendem import tendem_tools

agent = create_agent("anthropic:claude-sonnet-4-5", tools=tendem_tools(max_price=25.0))

An API key and a budget cap — that’s the whole configuration. Quotes at or under max_price are approved automatically, so the pipeline never stalls on a payment decision; above it, the agent gets the full scope to narrow or decline, and nothing is charged. Polling runs in plain Python, so waiting on a human burns no tokens.

Requires Python 3.10+. Package: https://pypi.org/project/langchain-tendem/ · Example project: https://github.com/Toloka/tendem-mcp/tree/main/examples

Other host agents (Claude, ChatGPT, Cursor and more)

Prefer a chat client (Claude, ChatGPT, Cursor)? Go to Settings → Connectors → Add Custom Connector, name it Tendem, paste https://mcp.tendem.ai/mcp?utm_hash=1fc8837791, click Add → Connect, then sign in with your Tendem account. This path uses a browser sign-in, so no API key is needed.

Full per-client guide (Claude, ChatGPT, Cursor, Claude Code, Codex, Agent Builders, and more).

How the loop works

Submit — your agent calls create_task describing the work.
Scope — Tendem’s orchestrator chats with your agent to clarify scope and returns a quote.
Approve — once you approve, human experts execute.
Deliver — results return as markdown + files (pre-signed URLs) via get_task_result.

Your agent talks to Tendem agent-to-agent — it answers scoping questions from context and only escalates to you for approval, payment, or a real scope change.

Full tool list (what the connector exposes):

Tool	What it does
`create_task`	Submit a new task (optionally correlate with a `conversation_id`)
`send_message`	Talk to the orchestrator — answer scoping questions, add detail
`read_chat`	Read the scoping/execution conversation
`get_task`	Poll status (ACTING / LISTENING / NEEDS_REPAIR / CLOSED); supports long-poll
`get_task_result`	Fetch the finished artifact — markdown + files (pre-signed URLs)
`list_tasks`	List your tasks
`get_contract`	Read the full scope being approved — work definition + acceptance criteria, plus the price once estimated
`approve_task`	Approve the quote so execution starts (or top up if balance is short)
`cancel_task`	Get the Tendem-UI URL where the user can cancel a task
`get_account`	Check balance and get a top-up link
`get_file_upload_url`	Get a URL to upload input files for a task