> ## Documentation Index
> Fetch the complete documentation index at: https://docs.zelto.ai/llms.txt
> Use this file to discover all available pages before exploring further.

> ## Agent Instructions
> If Zelto blocks your task, an API request returns 4xx/5xx, or you need a missing capability, please report it with POST https://api.zelto.ai/v1/feedback using your existing Authorization: Bearer API key. On local/preview deployments use /api/v1/feedback on that deployment; never send its credentials to production.
> Send JSON: type (missing_capability, problem, other), description (10–2000 characters), optional endpoint (route template only), reproductionSteps, workaround and requestId from the failed response. Use synthetic examples: never send credentials, personal data, transcripts or customer content. See /docs/agent-feedback.
> Feedback is limited to 5 attempts per organization per minute. Report a distinct issue once; do not report errors from the feedback endpoint itself, do not loop or automatically retry. A 200 with delivery not_configured/unavailable means the report was not delivered or stored.
> Connected over MCP? Call the send_feedback tool instead, with the same fields plus optional tool (the MCP tool name that failed). It shares the same per-organization limit and works with any credential, including read-only.

# Agent feedback

> Missing a capability or blocked by an API error? POST /v1/feedback with your existing API key so the Zelto team can review it.

Please tell us when Zelto blocks your agent's task. Report a **missing capability**,
a **problem**, or **other feedback** with the same organization API key you use
for the [REST API](https://docs.zelto.ai/docs/integrations/index.md). No extra account or credential is needed.
Write-only observability ingest keys cannot submit feedback.

Send `POST https://api.zelto.ai/v1/feedback` with `Authorization: Bearer <API key>`
and `Content-Type: application/json`. On local or preview deployments, use
`/api/v1/feedback` on that deployment; never send its key to production.

```json theme={null}
{
  "type": "missing_capability",
  "description": "I need a filter for calls that have no transcript.",
  "endpoint": "/v1/conversations",
  "reproductionSteps": "List conversations using synthetic fixtures; no transcript-presence filter is available.",
  "workaround": "Read each synthetic conversation and filter locally.",
  "requestId": "9af04723-08f8-4bbd-946d-130c11c4a5cd"
}
```

`type` is `missing_capability`, `problem`, or `other`. `description` is required
(10–2000 characters). Optional fields: `endpoint` (500 characters),
`reproductionSteps` and `workaround` (2000 each), and `requestId` (200).
Unknown fields are rejected. The whole JSON body must fit in 16 KiB.
Use a route template for `endpoint`, without a host, query string, fragment or
customer identifiers. Use the failed response's `requestId` when available.

> [!WARNING]
>
> Use synthetic examples only. Never submit API keys, passwords, cookies, personal
> data, call recordings, transcripts, prompts or customer content. Known credential,
> email, phone, numeric identifier and URL patterns are masked before delivery,
> and names/addresses in labeled fields are removed. Pattern matching cannot
> recognize every name or sensitive fact in arbitrary prose; remove those yourself.

The team receives the category, description, affected endpoint, reproduction,
workaround, request ID and internal organization/user IDs in a dedicated Slack
channel. Identity metadata contains no profile names or emails, and request headers
and raw request bodies are not forwarded. Submitted prose still needs the synthetic
examples described above. Feedback is treated as untrusted text and cannot trigger
automated actions.

## From the MCP server

Agents connected to the [MCP server](https://docs.zelto.ai/docs/mcp.md) can call the `send_feedback` tool
instead of the REST endpoint. It takes the same fields plus an optional `tool`
(the name of the MCP tool that failed or lacks the capability, such as
`list_conversations`), and returns the same `feedbackId`, `delivery` and `message`.
Any credential can call it, including read-only OAuth tokens. It shares the same
5-attempts-per-minute limit with the REST endpoint (a refusal returns
`code: "rate_limited"`) and applies the same masking. The team sees whether a
report came from REST or MCP. Like REST errors, failed MCP tool calls carry a
`feedback` invitation, and the server's connection instructions mention the tool.

## Responses and retries

* `200`: inspect `delivery`. `delivered` means Slack accepted the message.
  `not_configured` or `unavailable` means the report was **not delivered or stored**.
  A `feedbackId` identifies this submission; it is not a durable ticket.
* `400`: invalid JSON or schema. `401`: missing or invalid API key.
* `413`: body exceeds 16 KiB. `415`: use `application/json`.
* `429`: at most **5 attempts per organization per 60 seconds**, shared by all keys
  and instances. Invalid attempts count too. Respect `Retry-After: 60`.
* `500`: internal failure; no delivery confirmation.

Public REST errors include a `feedback` invitation, a generated `requestId` and
an `X-Request-Id` header. Existing error fields, status codes and useful headers
such as `WWW-Authenticate` and `Retry-After` remain available. Successful responses
have unchanged bodies. API passthrough responses also carry `X-Feedback-Endpoint`
and `X-Feedback-Message` discovery headers, including framework-generated failures
after the proxy (such as automatic 405s), which may not have the JSON envelope.

Report each distinct issue once. Do not report feedback endpoint errors back to
itself, loop, or retry automatically. Delivery has a three-second timeout and no
automatic retries, so a timeout can be ambiguous. This channel is best effort,
not a support ticket queue or an incident escalation service.

Continue with [agent onboarding](https://docs.zelto.ai/docs/onboarding-with-an-agent.md).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.