Skip to Content
Builder Docs
CRM, APIs & Knowledge Bases

MCP & Third-Party Integrations

Connect external Model Context Protocol (MCP) servers over Streamable HTTP with tool allowlists, namespacing, and Talkar's hosted inbound `/api/v1/mcp` server.

MCP & Third-Party Integrations

Model Context Protocol (MCP) lets you expose an entire suite of external tools (such as Google Calendar, Notion, Linear, Zapier/n8n MCP bridges, or custom company servers) to a Talkar voice or chat agent without defining every single REST endpoint by hand.


1. Connecting Outbound Remote MCP Servers

In Talkar's Tools system, choose tool category mcp to register a remote MCP server.

Transport & Timeout Configuration

SettingDefaultPurpose
transport"streamable_http"Talkar connects to remote MCP servers over Streamable HTTP.
url(Required)Remote MCP endpoint URL.
credential_uuidOptionalReference to a Credentials Vault entry (including oauth2_client_credentials).
tools_filter[] (all)Optional allowlist of raw remote tool names to expose to the LLM. Leave empty to expose all discovered tools.
timeout_secs30HTTP connection & protocol handshake timeout.
sse_read_timeout_secs300SSE stream read timeout. Individual tool calls enforce sse_read_timeout_secs + 5 seconds.

Discovery, Caching & Refreshing Tools

  • On Save: Creating or updating an mcp tool performs a best-effort discovery probe against the remote server. If the server is temporarily unreachable while saving, Talkar saves the tool with an empty cache instead of blocking creation.
  • Explicit Refresh (POST /api/v1/tools/{tool_uuid}/mcp/refresh): Probes the remote MCP server on demand and updates the cached discovered_tools list. If the probe fails, your previously cached tool list stays intact while the endpoint returns the error reason.

How Tools Appear to the Agent (mcp__<slug>__<name>)

To prevent tool-name collisions across multiple connected servers, Talkar automatically namespaces every discovered MCP tool before passing it to the LLM:

mcp__<slug>__<remote_tool_name>
  • <slug> is normalized from your Talkar tool's name (falling back to the first 8 characters of tool_uuid or "server").
  • Allowlist filtering (tools_filter vs node mcp_tool_filters): Set raw remote tool names in the tool's tools_filter (leave empty to expose all discovered tools), or override per Agent node via mcp_tool_filters (null exposes all allowed tools; explicit empty list [] suppresses all tools from that server on that node). Exposing dozens of unused MCP tools bloats the LLM context and slows voice latency.

Resilience & Auth Refresh at Call Time

  • Graceful degradation: If the external MCP server is down when a call starts, the session degrades cleanly without crashing the voice conversation.
  • Automatic OAuth2 401 retry: If an MCP handshake or tool call fails with 401 Unauthorized while using an oauth2_client_credentials credential, Talkar automatically invalidates the cached OAuth2 token, fetches a fresh token, reconnects once, and retries the tool invocation.

2. Using n8n.talkar.in as Your Integration Hub

Not every SaaS app has a clean Streamable HTTP MCP server—and many third-party MCP servers return verbose nested JSON objects that take the voice LLM too long to parse.

For Google Calendar, Google Sheets, WhatsApp/Slack alerts, Zoho, HubSpot, or Salesforce:

  • Option A (Fastest voice path): Put an n8n webhook on https://n8n.talkar.in, do the multi-step CRM/calendar work inside n8n, return 3–5 clean fields, and call it from a normal http_api tool.
  • Option B (Multi-tool MCP): Expose an n8n MCP trigger over Streamable HTTP and filter only the 2–4 tools your agent requires via tools_filter.

3. Inbound Talkar MCP Server (/api/v1/mcp)

Beyond calling outbound MCP servers, Talkar also mounts its own hosted MCP server over stateless Streamable HTTP at:

/api/v1/mcp
  • Authenticates external MCP clients (such as IDE coding assistants or internal orchestration agents) using the exact same X-API-Key header as the REST API.

On this page