Context and tools
Give the bot documents to read and MCP servers to act on during the call
A prompt tells the bot who it is. Two more fields tell it what it knows and what it can do.
Documents and URLs: prompt_data_sources
Up to ten sources are loaded when the bot starts and appended to its system prompt as labelled context blocks:
{
"meeting_url": "...",
"prompt": "You are Analyst. Answer questions about the account using the notes.",
"prompt_data_sources": [
{ "name": "Account notes", "type": "url", "url": "https://example.com/acme-notes.md" },
{ "name": "Pricing", "type": "text", "text": "Starter is $49/month, Team is $199/month..." }
],
"prompt_data_token_limit": 6000
}| Field | Notes |
|---|---|
name | Label shown to the model above the block |
type | text (inline, in text) or url (fetched, from url) |
headers | Optional request headers for url sources, e.g. a bearer token; only over https:// |
token_limit | Per-source cap, applied before the request-level cap |
prompt_data_token_limit | Total cap across all sources, default 4000, max 50000; 0 disables loading |
Sources are fetched once, at bot creation, with a 1 MB limit per URL. URLs on private networks or localhost are refused on the hosted API. A source that cannot be used fails the request instead of producing a bot that silently knows nothing: 400 for a bad source (refused host, too large, bad type), 502 when the URL could not be fetched.
Keep the budget in mind: everything you attach is sent to the model on every turn. A few thousand tokens of focused notes work better than a whole document.
Tools: mcp
The bot can call tools on your Model Context Protocol servers mid-conversation — look up a customer, check a calendar, create a ticket — and speak the result.
{
"meeting_url": "...",
"prompt": "You are the support bot. Use the CRM tools to answer account questions.",
"mcp": {
"instructions": "Only read data. Never modify an account during the call.",
"servers": [
{
"name": "crm",
"transport": "streamable_http",
"url": "https://mcp.example.com/mcp",
"headers": { "Authorization": "Bearer ..." },
"tool_allowlist": ["get_account", "list_recent_calls"],
"timeout_seconds": 20,
"instructions": "Accounts are keyed by company name."
}
]
}
}| Field | Notes |
|---|---|
transport | http, streamable_http or sse. Omit it for a metadata-only server: the bot is told about the tools but cannot call them |
url, headers | Where to connect and how to authenticate. Headers only over https:// (or a private URL on a self-hosted server) |
tool_allowlist | The only tools the bot may call from this server; everything else the server exposes is hidden |
tools | Tool names to mention to the model when the server is metadata-only |
timeout_seconds | Per-server connection and tool-call timeout (0.1–300) |
instructions | Per-server guidance; mcp.instructions applies to all servers |
enabled | false documents a server without connecting it |
Up to ten servers per bot. Tool calls run one at a time while the bot holds the floor, so keep tools fast; anything over a few seconds is dead air in the call. Headers you pass are kept in memory for the bot's lifetime and never written to disk or returned.
Stdio MCP servers are not available through the API. The mcp_profile presets (professional, personal, all) connect to a local mcpproxy and only make sense on a self-hosted server that runs one; on the hosted API they fail to connect.