Skip to main content
If you already run an AI coding agent — Claude Code, Cursor, Codex, or similar — it can do most of the Zelto setup itself. The shortest path is one sentence: install zelto.ai. The agent reuses configured access or guides you through browser sign-in and OAuth authorization when needed, wires your code to send conversations, and checks that the first one arrived. From there it can keep going over the MCP server: create a monitor, wait for the first finding, and set up delivery. This page is both the human walkthrough and the script the agent follows (it reads the same steps through the read_docs tool).

The minimal prompt

From the voice agent project’s root, tell your coding agent:
The agent should discover this guide through https://zelto.ai/llms.txt and follow https://dashboard.zelto.ai/skill.md. Inspect the project and reuse any existing ZELTO_API_KEY, endpoint configuration or authenticated Zelto connection first. Do not register a second account or require MCP when an API key already works. Install the runtime integration at the completed-conversation hook, then send a clearly labeled synthetic conversation and read its transcript back. A receipt alone is not verification. Findings and monitors are optional after installation. With existing credentials this can run without manual steps. A new account still requires the owner’s email verification and agreement to the Terms and Privacy Policy. The agent must never bypass these or invent an account identity.

Install Zelto with one sentence

Paste this into your coding agent (Claude Code, Cursor, Codex, …) from the root of the project you want to observe:
Paste into your coding agent

Reach a first visible result on Free

After a real conversation arrives, ask your own coding agent to read its transcript, identify one concrete issue with supporting evidence and suggest a next action. If there is no substantiated issue, review another conversation instead of inventing one. Ask it to publish with create_finding only after you approve the proposal, linking the conversation as evidence. The agent should read the finding back with get_finding and give you the dashboard link. This does not require a paid monitor. The guide at the app’s /agent-setup.md walks through this after connecting calls, even when setup is already complete. Get started keeps the copyable prompt and a View findings link available after setup. Before publishing, the agent checks existing findings to avoid duplicates. The agent gives you Sign in or Create an account. Enter your credentials yourself and click Verify my account in your email. The email code remains available as a manual alternative in Zelto; never send it to the agent. Links and codes expire together after ten minutes. Connect your MCP client to https://api.zelto.ai/mcp using its normal OAuth flow, then approve organization access in your browser. The client manages PKCE, state and tokens. Existing accounts use the same flow, including social sign-in and MFA. A browser session is never handed to the agent. For discovery, the public https://api.zelto.ai/mcp/signup endpoint exposes start_signup with {}. It returns sign-in, registration and MCP URLs only. It does not create accounts, generate passwords, or request email codes. For server-side ingestion, create an API key in Settings > Integrations > API integration and configure it directly in your local env/secret store. Clients without OAuth can use that key locally too. Never paste it into chat. Then follow the conversation upload guide and continue to the first finding below. If you open the verification link in another browser or for a different account, Zelto shows the email address and asks you to confirm before signing you in.

Already signed in? One line from Get started

If you already have a Zelto account, open Get started in the dashboard. A new organization opens on Set up Zelto with your coding agent; after you skip it, the same line stays under Or let your coding agent set it up, below the steps. Copy it into your agent. It looks like this:
The guide at /agent-setup.md has the agent add the MCP server to itself and sign in with OAuth, so the line holds no API key: you sign in once in your browser and pick the organization. The agent then reads get_onboarding_status and walks the flow below with you: it connects your voice platform or wires your own code to the ingest API, and offers a first monitor. It never signs up or creates another organization.

Before you start (if you already have a key)

If you already created a key in the dashboard, skip straight to connecting the MCP server:
  1. Have an API key. In the dashboard it is at Settings → Integrations → API integration, shown once when you create it.
  2. Connect the Zelto MCP server in your editor with that key — see MCP for the per-client config. Confirm with the health tool (it returns your org id).
Everything below runs against the onboarding tools on that connection.

The flow

1

Check where you are

The agent calls get_onboarding_status first. It returns the ordered setup steps — connectCalls (conversations are arriving), importAgent (an agent exists), createMonitor — each with a done flag, plus nextIncomplete and monitorsAvailable. Re-running onboarding is safe: the agent resumes from the first incomplete step. If you sent conversations with the conversation upload guide, the first two are already done.
2

Connect a conversation provider

The agent calls list_integrations, asks which stack you run, then calls connect_provider with the provider and an optional account label. It returns a link valid for 15 minutes. Open it as an owner or admin of your organization and enter the key directly in Zelto; you never paste it into the agent chat. Zelto validates the key and stores it encrypted. The agent checks check_provider_connection with the returned token before importing agents.For your own AI-model or speech-to-text keys, the agent uses connect_ai_provider and the same browser flow and completion check. Custom OpenAI-compatible endpoints also collect a base URL and model IDs in the web form; validation incurs a small provider charge per model.Optional voice-provider webhook instructions remain in Settings → Integrations.
3

Get conversations flowing

Conversations arrive once the webhook is pasted (or the hourly sync runs for the pull-based providers). The agent polls list_conversations until the first conversation lands. On a first conversation, Zelto creates the agent automatically. Already have history in a connected Retell account? The agent can import it with retell_import_agent and retell_import_conversations instead of waiting.If this integration will run experiments, also read Agent versions. LiveKit and custom uploads should report an explicit top-level version on every conversation. Verify conversations from two release labels appear under the same agent before configuring a comparison.
4

Create a monitor

A monitor is an AI judge that runs against every conversation and flags the ones that match a rubric. Monitors are available on every plan; estimate usage costs before creating one. The agent offers the ready-made catalog from list_starter_monitors (refusals, task failure, user frustration, and more), then calls create_monitor — either with a starterKey, or with a name and a plain-language yes/no instruction of your own. It confirms with list_monitors.
5

Wait for the first finding

As conversations flow through the monitor, Zelto surfaces findings — durable, org-level issues grouped from the conversations that matched. The agent polls list_findings until the first one appears. The agent can also create a finding itself with create_finding (a plain-string description is fine); it is open and visible immediately, including on Free.
6

Route findings to your team

Finally the agent calls set_findings_delivery with the channels you want — email and/or slack. Slack additionally needs the Slack app connected via OAuth under Settings → Integrations, which the agent will flag for you to do.

Use your coding agent

Paste the prompt below into the coding agent in your editor. It loads the zelto-onboarding skill if your client supports Anthropic Agent Skills, or falls back to reading this page through read_docs — either way it follows the flow above and pauses for the steps only you can do.
Paste into your coding agent

Install the skill

Three ways to give your agent the skill, lowest-friction first:
  1. Zero-install (works with any MCP client). Every connected editor already has the read_docs tool — the paste-prompt above tells the agent to read_docs('onboarding-with-an-agent') and follow it. This always reflects the live version of this page. Recommended.
  2. Copy the skill folder (Claude Code and other Skills-capable clients). Drop the zelto-onboarding skill folder into your agent’s skills directory (for Claude Code, .claude/skills/). The skill’s SKILL.md plus its references/ files drive the same flow with progressive detail.
  3. Reference it inline. Point the agent at this page’s URL in your own prompt; it will read_docs the current version on demand.
Blocked by an API error or missing a capability? Send agent feedback with your existing API key.