The Qonversion MCP server is currently in beta. Tool names, arguments, and OAuth scopes may still change. Not recommended for production-critical workflows yet — we’d love your feedback while we stabilise it.
https://mcp.qonversion.io/mcp and uses OAuth 2.1 for authentication.
The server follows the MCP Streamable HTTP spec and supports clients that connect natively over HTTP, as well as clients that use mcp-remote for compatibility.
The MCP server mirrors the Qonversion REST API v4 resource surface — each MCP tool corresponds to a v4 operation. Authentication is different, though: HTTP clients authenticate with the project Secret Key, while MCP clients use OAuth 2.1 against your Qonversion account with per-tool scopes. See the full v4 guide for resource-by-resource context, or the API Reference for the underlying HTTP operations.
Capabilities
The Qonversion MCP server exposes the full surface of the v4 REST API as MCP tools, grouped by resource:- Customers — list, read, delete (GDPR), set properties, list and mutate permissions, and read aggregated metrics.
- Users, identities, purchases, properties — look up users, attach external identities, inspect purchase history, set and query user properties.
- Entitlements — manage entitlement definitions and grant or revoke them per user.
- Products & Offerings — CRUD products and configure offerings consumed by the SDK.
- Paywalls (Screens) — list, create, update, publish, duplicate, and pull per-screen and overview analytics.
- Experiments — CRUD experiments and groups, change status, attach and detach users, read summary analytics.
- Segments — CRUD user segments and list system segments.
- Analytics — charts, KPI cards, cohorts, LTV, trial conversion, currencies, and narrative insights.
- Exports — start asynchronous data exports and poll their history and status.
- Events — inspect the event catalog and the raw event log.
- Scheduled Reports — CRUD recurring reports, list destinations, send a test delivery.
- Integrations — discover the catalog, create, toggle status, and delete integrations.
- Automations — CRUD event-driven automations and change their status.
- Project settings — read and update project configuration, rotate the Secret Key, manage store credentials.
customers:read, screens:read, experiments:write). Your effective permissions mirror your role in the Qonversion dashboard.
Destructive actions
Destructive tools (delete_*, revoke_*, detach_*, regenerate_project_secret) use a two-step confirmation flow designed around human-in-the-loop intent. The confirmation token is addressed to the operator, not to the agent — the agent is expected to surface the pending action and wait for explicit approval before executing.
The flow
- The agent calls the destructive tool (e.g.,
delete_product(product_id: "prem_monthly")). - The server does not perform the action. It returns a message listing the pending action, the target Resource id, and a short-lived token (5-minute TTL). The message is explicitly addressed to the human operator.
- The agent must surface the pending action to the human and wait for explicit approval.
- Once approved, the agent calls
confirm_destructive_actionwith three required arguments:
| Field | Type | Notes |
|---|---|---|
token | string | The opaque token from step 2 |
resource_id_echo | string | Must match the Resource id verbatim (case-sensitive). A mismatch cancels the pending action and returns a typed error. |
reason | string | Short justification, 1-256 characters. Captured in the audit log. |
Why the echo + reason
These two fields turn a single-turndelete → confirm chain into an act of deliberate re-statement. The agent can’t forward the token blindly — it has to restate the target (which surfaces the action to the human) and commit a reason (which is easy to spot in review if fabricated). Silent auto-confirmation stops being possible without the model making a distinct act that’s visible in the transcript.
Tokens
- Tokens are bound to the MCP session that requested them — another session can’t consume them.
- Tokens expire after 5 minutes. Expired tokens must be re-requested from the original tool.
- A token can be used at most once. A mismatched
resource_id_echoconsumes the token and cancels the pending action. - Tokens are not logged in the audit trail; the resource id and reason are.
Destructive tools affected
delete_automation, delete_customer, delete_entitlement, delete_experiment, delete_experiment_group, delete_integration, delete_offering, delete_product, delete_remote_configuration, delete_scheduled_report, delete_screen, delete_segment, detach_experiment_user, detach_user_from_remote_configuration, revoke_customer_permission, revoke_user_entitlement, regenerate_project_secret.
Setup
Claude Desktop
Add the following to yourclaude_desktop_config.json:
Claude Code
/mcp in a Claude Code session to go through the authentication flow.
Cursor
Add the following to your.cursor/mcp.json:
Visual Studio Code
- Press
CTRL/CMD + Pand search for MCP: Add Server. - Select Command (stdio).
- Enter the following and press enter:
- Enter the name Qonversion and press enter.
- Open MCP: List Servers, select Qonversion, and click Start Server.
Windsurf
- Press
CTRL/CMD + ,to open settings. - Go to Cascade → MCP servers.
- Select Add Server → Add custom server and add the following:
Zed
- Press
CMD + ,to open settings. - Add the following:
Codex
Add the following to your~/.codex/config.toml:
codex mcp login Qonversion to authenticate.
Other clients
For clients that support MCP via stdio, use the following configuration:- Command:
npx - Arguments:
-y mcp-remote https://mcp.qonversion.io/mcp - Environment: None