Integrations
An integration is an account you have connected — your Gmail mailbox, your Xero organisation, an MCP server. Connect one once, and your agents can act in it.
Connectors and integrations
Two words that sound similar and mean different things:
- A connector is a type of account Tunori knows how to talk to, such as Google Drive. Every connector is listed in the connector reference.
- An integration is your connected account of that type. You can have several integrations of the same connector — two mailboxes, two Xero organisations — and name them to tell them apart.
Integrations belong to the organization, not to one person or one agent.
Connecting an account
From Integrations, choose Add integration:
- Pick a connector. Search the catalog for what you need.
- Name it. Use something you will recognise later, especially with more than one account of the same kind.
- Sign in with the provider and choose what to grant.
Connecting an account requires the Admin role. See teams and roles.
Permissions
You decide what to grant on the provider's consent screen, and Tunori works with whatever you allow. Granting less is a valid choice — it simply means fewer tools are available to your agents.
Each integration shows Granted permissions, and lists its tools and triggers split into available and unavailable, so you can see exactly what a narrower grant costs you. If something you need is unavailable, reauthorize the integration and grant the missing permission.
Giving agents access
Connecting an account does not automatically give every agent access to it. In each agent's settings you enable the integrations that agent may use, choose which of its tools are enabled, and mark the tools that require your approval before running. See agents.
MCP servers
MCP — the Model Context Protocol — is an open standard for connecting AI to software. Rather than supporting one fixed set of tools, Tunori can point at any MCP server: your issue tracker, your CRM, your own internal tooling.
Point Tunori at the server's URL and its tools, resources and prompts are discovered automatically and become things you can ask for in plain language. Authenticate with OAuth or a bearer token; credentials are encrypted and never shown to the agent.
Because the tools come from the server, they are not listed in this documentation — you see them once the server is connected. See the MCP connector.
HTTP requests
Not every system speaks MCP. The HTTP Request connector lets you call any HTTP API or webhook: an administrator configures one request per integration — the URL, method, headers and the kind of payload it takes — together with a short description of what it does. The agent then has a single tool for it.
Anything the agent should fill in goes into {{placeholders}} in the URL, header values or body,
for example https://api.example.com/orders/{{orderId}}; each placeholder becomes a required input
of the tool. For requests with a payload you can write the body yourself, e.g.
{"name": "{{name}}"} — the agent then only fills in its placeholders (values are escaped for JSON
or form bodies). Leave the body empty to let the agent compose the whole payload.
Header values are never shown to the agent. Mark a header as secret for API keys and tokens: its value is encrypted and write-only — when you edit the integration, leave it blank to keep it. Other header values stay visible in the edit form.
The agent receives the status, headers and body of the response. Text bodies come back inline (up to 100 KB); the agent can also save the full body — up to 10 MB, which is the only way to obtain binary responses such as PDFs or images — into its files. Requests time out after 30 seconds, and only public hosts can be called: private and internal network addresses are refused. See the HTTP Request connector.
When something stops working
Integrations can end up disconnected or in error, usually because access was revoked at the provider or a token expired. The integration shows its status and why it stopped working — access expired or was revoked, or the provider asks for a fresh sign-in — and reauthorizing restores it. Triggers and schedules on an inactive integration are skipped rather than failing silently.