Add Tendem integration

integrations/tendem.md

@@ -0,0 +1,130 @@
+---
+layout: integration
+name: Tendem
+description: Delegate tasks to vetted human experts - research, writing, analysis, and data work.
+authors:
+    - name: Tendem
+      socials:
+        github: Toloka
+repo: https://github.com/Toloka/tendem-mcp
+report_issue: https://github.com/Toloka/tendem-mcp/issues
+type: Tool Integration
+version: Haystack 2.0
+toc: true
+---
+### **Table of Contents**
+- [Overview](#overview)
+- [Installation](#installation)
+- [Usage](#usage)
+- [License](#license)
+
+## Overview
+
+[Tendem](https://tendem.ai) is a hybrid AI + human task service, exposed to Haystack as a
+remote MCP server. Your agent submits a task in natural language; Tendem's orchestrator
+scopes it and quotes a transparent price; after explicit approval a vetted human expert
+executes it and returns verified results as markdown plus files.
+
+This gives a Haystack `Agent` a way to hand off the work that a model alone cannot reliably
+finish — research and competitive analysis, fact-checking, copywriting and editing, design
+review, presentation polish, data cleaning and list building. Tendem declines data scraping
+and extraction tasks by policy, and quick general-knowledge questions are better answered by
+the model itself.
+
+The server is hosted at `https://mcp.tendem.ai/mcp` over streamable HTTP, so there is no
+Tendem-specific Haystack package to install — the
+[MCP integration](https://haystack.deepset.ai/integrations/mcp) connects to it directly.
+
+The main tools it exposes are `create_task`, `send_message` and `read_chat` (agent-to-agent
+scoping), `get_contract` (scope and price), `approve_task`, `get_task` and `get_task_result`.
+
+> Tasks are billed. `create_task` and scoping are free; nothing is charged until
+> `approve_task` is called, so keep a human in the loop before approving a quote.
+
+## Installation
+
+```bash
+pip install mcp-haystack
+```
+
+## Usage
+
+Interactive clients authenticate with OAuth 2.0 on first use. For scripts and pipelines,
+create an API key at [agent.tendem.ai/tokens](https://agent.tendem.ai/tokens) and pass it as
+an `Authorization` header:
+
+```bash
+export TENDEM_AUTH="ApiKey <your-token>"
+```
+
+### Load the Tendem tools
+
+```python
+from haystack.utils import Secret
+from haystack_integrations.tools.mcp import MCPToolset, StreamableHttpServerInfo
+
+toolset = MCPToolset(
+    server_info=StreamableHttpServerInfo(
+        url="https://mcp.tendem.ai/mcp?utm_hash=b3b2da7bd9",
+        headers={"Authorization": Secret.from_env_var("TENDEM_AUTH")},
+    ),
+)
+
+for tool in toolset:
+    print(tool.name)
+```
+
+### Use it in an Agent
+
+```python
+from haystack.components.agents import Agent
+from haystack.components.generators.chat import OpenAIChatGenerator
+from haystack.dataclasses import ChatMessage
+from haystack.utils import Secret
+from haystack_integrations.tools.mcp import MCPToolset, StreamableHttpServerInfo
+
+toolset = MCPToolset(
+    server_info=StreamableHttpServerInfo(
+        url="https://mcp.tendem.ai/mcp?utm_hash=b3b2da7bd9",
+        headers={"Authorization": Secret.from_env_var("TENDEM_AUTH")},
+    ),
+    tool_names=["create_task", "read_chat", "send_message", "get_contract", "get_task"],
+)
+
+agent = Agent(
+    chat_generator=OpenAIChatGenerator(model="gpt-4o-mini"),
+    tools=list(toolset),
+    system_prompt=(
+        "You can delegate work to human experts through Tendem. Pass the user's own words "
+        "into create_task and let Tendem drive scoping. Always call get_contract and show "
+        "the scope and price to the user, and never call approve_task without an explicit "
+        "go-ahead. Poll with get_task(task_id, wait_for_change_seconds=30) rather than "
+        "looping."
+    ),
+)
+
+result = agent.run(
+    messages=[
+        ChatMessage.from_user(
+            "Have a human expert compare the top 5 open-source vector databases "
+            "on licensing and hybrid search support, with sources."
+        )
+    ]
+)
+print(result["messages"][-1].text)
+```
+
+Selecting a subset of tools with `tool_names` keeps the approval and cancellation tools out
+of the model's reach, which is a good default for unattended pipelines.
+
+### Retrieving results
+
+Once a task is approved and executed, `get_task_result` returns markdown plus pre-signed URLs
+for any files the expert produced, so results can be fed straight into a Haystack pipeline —
+for example converted and written to a Document Store.
+
+## License
+
+The Tendem MCP server is a hosted commercial service; use is governed by the
+[Tendem terms](https://tendem.ai). The client-side plugin configuration in
+[Toloka/tendem-mcp](https://github.com/Toloka/tendem-mcp) is Apache-2.0 licensed.