Skip to main content

MCP servers

Model Context Protocol (MCP) is an emerging standard for connecting AI assistants to tools and data sources. Instead of registering each function one by one in Insighto, you point Insighto at an MCP server URL — and the assistant automatically discovers every tool that server advertises, with their parameter schemas, on the next conversation. Add a tool to the server tomorrow and your assistant has it tomorrow, no Insighto-side change needed.

If you're building a tool surface that's going to be reused across Claude, ChatGPT, IDE assistants, and Insighto, MCP is the right shape. If you just need to call one endpoint, a Custom tool is simpler.

What you can do with an MCP server​

  • Point Insighto at any MCP server URL and have the assistant auto-discover its tools.
  • Update the tool list on your side without touching Insighto — added/removed tools show up on the next conversation.
  • Use the same MCP server with multiple AI products — write once, integrate everywhere.

Business use cases​

Expose your internal services to the assistant via MCP​

A B2B SaaS company runs several internal services — billing, provisioning, support history. They build one MCP server in front of all of them with tools like get_account_status, provision_seat, lookup_support_history. Their voice agent, their internal Claude-based admin tool, and their ChatGPT-powered onboarding bot all use the same MCP server. One implementation, three product surfaces.

Use a partner's MCP server​

A logistics platform publishes an MCP server for tracking and dispatch operations. Customers connect that server to their own AI workflows. Insighto users plug the URL into their account's connections and get tracking, dispatch creation, and route lookup as assistant tools without writing a single integration.

Iterate quickly on new tools​

A dev team is building a tool surface for a customer-success use case. They host their MCP server and add tools as they go — get_open_renewal_risk, summarize_recent_tickets, flag_for_QBR. Each new tool appears in their Insighto assistant on the next conversation. No new Insighto config required, no separate "register this function" step.

Curated tool surface for an agency's clients​

An agency builds a single MCP server that exposes the right set of tools for each client business type — dental practice tools, plumbing-business tools, real-estate brokerage tools — gated by URL path. Each client's Insighto assistant connects to a different URL on the same server and sees only the tools relevant to their vertical. Centralized maintenance, decentralized exposure.

What MCP servers expose​

MCP servers can expose tools, resources, and prompts. Insighto currently consumes only tools — the function-callable kind. Resources (read-only data sources) and prompts (parameterized templates) aren't part of this integration today. If your MCP server publishes all three, only the tools become available to the assistant.

How discovery works​

Discovery happens at the start of every conversation. If you add, remove, or change a tool on your server, the change shows up on the next conversation. There's no manual refresh button and no background cache.

Connecting an MCP server​

  1. Go to Tools & Integrations → MCP → Add MCP Server.
  2. Server URL. Paste your MCP server's endpoint — an https://-prefixed URL pointed at a remote MCP server that speaks SSE or HTTP transport. Stdio-based local MCP servers (the kind used by Claude Desktop) aren't usable here; those are for local desktop agents.
  3. Auth. Bake authentication into the URL itself — typically a signed-URL query parameter or a path-embedded token. There isn't a separate username/password field today, so any auth you need (bearer tokens, API keys) has to be enforced by the server at the URL level or by a reverse proxy in front of it.
  4. Save. Insighto stores the URL.
  5. Attach to assistants. Pick the MCP server from the assistant's Tools tab. The tools the server advertises become callable on the next conversation.

Auth — what to do today​

Until Insighto adds a first-class auth UI for MCP servers, the practical paths are:

  • URL-embedded credentials — a signed URL or token-as-query-param: https://mcp.example.com/v1?token=abc123.
  • Reverse proxy with header injection — front your MCP server with a proxy that authenticates incoming requests via some shared secret and injects the credentials needed by the upstream server.
  • IP allowlist — limit access to Insighto's egress IPs.

Don't ship a public unauthenticated MCP server unless every tool on it is safe for anyone to call.

How an MCP tool looks to the assistant​

Exactly like any other tool. The assistant sees a function definition with a name, a description, and a parameter schema. It decides when to call the function based on the system prompt and the customer's input. Insighto handles the MCP protocol details transparently — the model just sees regular function calls.

This means the same prompting practices apply:

  • Mention the relevant tool name in the assistant's system prompt with a clear trigger.
  • Make sure the descriptions on your MCP server's tools are sharp — that's what the model reads. Vague descriptions equal unreliable invocation.

Example invocation​

Suppose your MCP server exposes a tool named search_tickets with schema:

{
  "type": "object",
  "properties": {
    "query": { "type": "string", "description": "Full-text search query" },
    "status": { "type": "string", "enum": ["open", "closed", "all"], "default": "open" }
  },
  "required": ["query"]
}

The assistant calls it:

{
  "function": "search_tickets",
  "arguments": {
    "query": "billing problem",
    "status": "open"
  }
}

Insighto forwards the call over MCP to your server, gets the response, and hands the result back to the assistant — which uses it to compose its reply to the customer.

System prompt guidance​

Even though tools come from the MCP server, the assistant still needs the prompt to know when to invoke them. Spell out triggers:

You have access to tools for looking up tickets. When the customer mentions a ticket number or a past issue, use search_tickets to find it. Read the ticket status and the latest update back to the customer in plain language.

The MCP server can have fifty tools registered, but if the prompt doesn't mention the relevant ones at the relevant trigger points, the model won't use them reliably.

MCP vs custom HTTP tool​

When to pick which:

  • MCP if you already speak the protocol, or want the same tool surface to work across Claude, ChatGPT, IDE assistants, and Insighto. Auto-discovery means you don't have to re-register functions in each tool surface.
  • Custom HTTP tool if you just want to call one HTTP endpoint and don't care about cross-platform compatibility. Simpler to set up and you keep the function schemas explicit inside Insighto.

For a brand-new integration being built from scratch, MCP is increasingly the better long-term choice. For wrapping a single existing endpoint, custom tools are faster.

Hosting an MCP server​

If you're building your own:

  • Use a well-maintained MCP framework. Official SDKs (TypeScript, Python) handle the JSON-RPC plumbing. Don't hand-roll the protocol.
  • Authenticate. See the auth section above.
  • Write good tool descriptions. Same advice as for custom tools — the description is what the model sees. Be specific about when each tool should be used and what each parameter means.
  • Host close to Insighto. Each MCP tool call is a network hop. For voice flows where latency matters, host the MCP server near Insighto's infrastructure.

Selective exposure​

Every tool the MCP server advertises gets registered with the assistant. There's no per-tool checkbox in Insighto to expose only a subset. Practical guidance:

  • Build narrow MCP servers — expose only the tools you want the assistant to use.
  • If you don't control the server, front it with a proxy MCP server that re-exposes a curated subset.

Failure modes​

  • Server unreachable on discovery — Insighto silently registers zero tools from that server and runs the conversation without them. The assistant won't have access to those tools but also won't fail loudly. Check your server's reachability and logs.
  • Auth expired or URL invalidated — same path as unreachable. Update the URL.
  • Tool removed from server — the assistant just doesn't see it on the next conversation. If your prompt referenced a function that no longer exists, the prompt should be tightened or the tool restored.
  • Schema drift — required parameters added or removed. Insighto re-pulls schemas every conversation, so the model sees the latest shape. Test after server upgrades.

Cost considerations​

MCP tool calls add a network hop on top of whatever your upstream service does. For voice flows where every 300ms of silence is noticed, host the MCP server geographically close to Insighto and keep tool execution time short.

Where to next​

  • For non-MCP services, build a Custom tool.
  • For built-in integrations (CRM, calendar, payments, support), see the Tools overview.
  • For prompt patterns that work across every kind of tool, see Assistants.