MCP server

This API serves a Model Context Protocol server over HTTP at https://api.salesbud.com.br/mcp. Registered once as a custom connector in claude.ai, it lets the model read meetings, calls, transcripts, template answers, evaluations, email and WhatsApp conversations as you or with a company credential — nothing to install.

It is a client of the public API, not a new surface. Same OAuth credential, same scopes, same rate limits, read operations only — nothing it can reach is anything the HTTP API would not return to the same credential.

Before you start

Your company needs the API_ACCESS feature enabled. What else you need depends on how the assistant signs in:

• Signing in with Salesbud (OAuth) — each person connects their own account and reads what they can see. Nothing to create beforehand. This is the way for Claude and ChatGPT: see Connect Claude or ChatGPT.
• A fixed credential in headers — an administrator enters one API credential that the whole Claude organization shares. You need that credential and owner access to the Claude organization. See the Quickstart for the two kinds of credential.

Sign in with Salesbud

Point the assistant at https://api.salesbud.com.br/mcp and let it sign you in: it finds the authorization server on its own, opens the Salesbud consent page, and connects as you. Step by step for each assistant, approval and disconnection are in Connect Claude or ChatGPT.

A connection made by someone who does not approve credentials waits for an administrator. Until then, every tool answers that the connection is awaiting administrator approval.

Use a fixed credential instead

An owner of the Claude organization can still add a custom connector pointing at https://api.salesbud.com.br/mcp with no sign-in, sending the credential of the company whose data the connector should read in two request headers:

Header	Value
X-Client-Id	sb_client_...
X-Client-Secret	sb_secret_...

Any other HTTP client may send the same credential as RFC 7617 Basic instead:

Authorization: Basic <base64 of client_id:client_secret>

Two consequences of a credential entered once by an administrator:

• It is stored on Anthropic’s side and shared by the whole Claude organization. Everyone in it reads whatever the scopes allow, and the audit trail records the credential, not the person who asked.
• Revoking means rotating. Rotating the secret breaks every other integration using the same credential, so give the connector one of its own.

If the credential restricts access by IP, the allowlist has to include Anthropic’s egress range, 160.79.104.0/21: the request reaches this API from there, not from the user’s browser.

Tools

One tool per public read operation. The scope column is what the credential has to carry — a missing scope comes back as a readable INSUFFICIENT_SCOPE message the model can report, not an exception.

Tool	Scope
get_api_context	none beyond a valid credential
list_meetings	meetings.read
find_meetings	meetings.read
get_meeting	meetings.read
get_meeting_overall_evaluation	meetings.read
get_meeting_answers	meetings.read
get_meeting_transcript	meetings.read + transcriptions.read
list_calls	calls.read
find_calls	calls.read
get_call	calls.read
get_call_overall_evaluation	calls.read
get_call_answers	calls.read
get_call_transcript	calls.read + transcriptions.read
list_emails	emails.read
get_email	emails.read
list_email_messages	emails.read + emails.content.read
list_whatsapp_conversations	whatsapp.read
get_whatsapp_conversation	whatsapp.read
list_whatsapp_messages	whatsapp.read + whatsapp.content.read
list_accounts	any of meetings.read, calls.read, emails.read, whatsapp.read
list_teams	any of meetings.read, calls.read, emails.read, whatsapp.read
list_tags	any of meetings.read, calls.read, emails.read, whatsapp.read

list_accounts searches by name or domain with q; list_teams and list_tags list the company’s teams and meeting and call tags. list_meetings and list_calls accept account_id, team_id, tag_id, score_min and score_max; list_emails and list_whatsapp_conversations accept account_id and team_id. A typical flow is list_accounts with q: "acme" to find the id, then list_meetings with that account_id. List results name every account of each record, up to three. See Filtering by account, team, tag and score.

Ask get_api_context first when you are unsure what a credential reaches: it answers with the client, whether it is a company or a user credential and, for a user credential, its owner, the company, the scopes and the per-minute limit.

/oauth/token is deliberately not a tool, and exposing it would burn the OAuth bucket and put the token in the conversation. With a fixed credential the server issues and renews the token itself. With a signed-in connection the assistant (Claude or ChatGPT) renews its own token with the refresh token. The liveness and readiness probes are not tools either.

The tool catalog

GET /mcp/tools lists the server’s tools without a credential: the same name, title, description, inputSchema and annotations that tools/list returns, plus the server version in serverVersion. A client can build its tool list, or notice that a description changed, without opening an MCP session. The answer may be cached for five minutes (Cache-Control: public, max-age=300). It describes the tools only; calling one still needs a credential.

curl https://api.salesbud.com.br/mcp/tools

Finding one account’s meetings and calls

“Show me the meetings with Acme” is one call to find_meetings (or find_calls). It resolves the account by name, domain, CNPJ or id in the account catalog and reads the records linked to it — the last 12 months unless you pass meeting_after or meeting_before. When none is linked, it reads the period record by record (90 days by default) and matches the account in the title, the linked accounts and the participants’ e-mail domain. A name that matches several accounts comes back with the candidates, so the model asks or picks by id instead of guessing.

One call reads at most 30 pages in 20 seconds. complete says whether the period was covered; when a ceiling stops the search, next_cursor continues it, passed back with the same account and filters. The answer carries the 50 most recent records, whole, in structuredContent, and only says an account has no records about a period it read to the end.

Transcripts arrive in windows

A one-hour transcript is thousands of utterances and does not belong in a model’s context in one piece. The transcript tools return 200 utterances per call and say how many are left; the model passes offset to continue. The API returns the transcript whole — this windowing lives in the MCP server.

A record with no transcript is not an error: the tool reports transcript unavailable (status: not_started), which is the contract described in Meetings and calls.

Email bodies cost more than email metadata

list_emails and get_email return conversation metadata — participants, connected mailboxes, linked accounts, counts — and never a body. Reading the messages is a separate tool, list_email_messages, and it asks for more: emails.content.read on top of emails.read, plus its own rate-limit policy, tighter than the rest of the API. Ask for a page at a time instead of walking a whole mailbox.

A body longer than 2,000 characters is trimmed in the text the model reads and kept whole in structuredContent. Attachment content, HTML bodies and bcc are never served by the API, so no tool reaches them.

WhatsApp text costs more than WhatsApp metadata

list_whatsapp_conversations and get_whatsapp_conversation return conversation metadata — contact, seller, linked accounts, activity instants — and never a message. Reading the messages is a separate tool, list_whatsapp_messages, and it asks for more: whatsapp.content.read on top of whatsapp.read, plus its own rate-limit policy. Ask for a page at a time instead of walking a whole history.

A text or an audio transcription longer than 2,000 characters is trimmed in the text the model reads and kept whole in structuredContent. A message deleted for everyone stays as a tombstone with no content; an edited message shows its current text. Media content, provider identifiers, reactions and read state never reach the model — the API does not serve them. See the WhatsApp reference.

What the server handles, and what it does not

It handles token renewal with a safety margin and one retry on 401 with a fresh token when it holds a fixed credential, Retry-After on 429, backoff on 503, and immediate failure on the other 4xx. With a signed-in connection the assistant renews its own token, and the server does not retry a 401: the tool answers with an error asking you to reconnect. Errors reach the model carrying code and request_id, so a bad date filter comes back as INVALID_DATETIME and the model corrects the format itself.

When the client cancels a call, or gives up on the request, the server stops the reads that call started. Every date parameter declares format: "date-time" in its input schema.

It does not paginate for you, except inside find_meetings and find_calls. The model follows has_more and next_cursor like any other client, and the tool description says so, because a short page is not the last page.

It also does not give you incremental sync, because v1 has none: window by meeting_at or created_at and reprocess idempotently by id.

Give a shared connector its own credential

A company can hold 20 active company credentials. Issue one for the connector alone, with only the scopes it needs, so revoking it on the API credentials screen never disturbs your other integrations. Use a company credential here, not a user credential: the connector is shared by the whole Claude organization, so everyone in it would read what the credential’s owner can see.

Shared quota

Every credential of a company draws from the same 600 requests/minute company bucket. An agent walking a long history consumes quota your other integrations are using; see Rate limits.