# VERANeural

> VERA is a personal companion that listens first, then helps people move: by chat or live voice, in the browser, on phones, on Meta Ray-Ban Display glasses, and inside other products through the VERA Partner API. VERA is not therapy, a clinician or emergency care.

## For developers

- [Developer overview](https://veraneural.com/developers): VERA Vibe Coding, VERA CLI, the Partner API and Connections.
- [Full documentation](https://veraneural.com/developers/docs): authentication, endpoints, safety levels, actions and approval, webhooks, errors, limits, security, changelog.
- [OpenAPI 3.1 specification](https://veraneural.com/developers/openapi.yaml): exact request and response shapes. Generate clients from it.
- [Postman collection](https://veraneural.com/developers/vera-partner-api.postman_collection.json)

## Partner API essentials

- Status: early access. Keys are issued after a security review and the signed VERA Partner Agreement. Pricing is agreed during onboarding.
- Base URL: https://api.veraneural.com/v1 . JSON over HTTPS only. Current version: 2026-10-09, returned in the VERA-Version header.
- Auth: `Authorization: Bearer vera_sk_...`. Keys stay on the server, never in a browser or mobile app.
- Scopes: conversations:read, conversations:write, actions:write, tools:write, webhooks:write, audit:read.
- Core loop: POST /v1/conversations, then POST /v1/conversations/{id}/messages. The reply has `message`, `safety` and `proposed_actions`.
- Safety: `safety.level` is none, elevated or crisis. When `safety.display_required` is true, show every resource in `safety.resources` prominently. Never hide, shorten or delay them.
- Actions: VERA proposes; the person approves. Call POST /v1/actions/{id}/approve only after the person explicitly agrees, then report POST /v1/actions/{id}/result. Actions expire after 24 hours.
- Tools: PUT /v1/tools/{name} with a JSON Schema `input_schema`. Up to 25 tools; up to 3 proposed actions per reply.
- Webhooks: HTTPS only, signed with `VERA-Signature: t=<unix>,v1=<hex HMAC-SHA256 of "t.rawBody">`. Reject timestamps older than 5 minutes. Retries after 1 s, 10 s and 60 s; deduplicate with `VERA-Event-Id`.
- Idempotency: send `Idempotency-Key` on POSTs and retry with the same key.
- Errors: `{"error": {"type", "message", "request_id"}}`. Types: invalid_request (400), authentication_error (401), permission_error (403), not_found (404), conflict (409), rate_limited (429, honour Retry-After), engine_unavailable (502, retry with the same Idempotency-Key).
- Limits: 60 requests per minute per key, messages up to 8,000 characters, context up to 4,000, request body 64 KB.
- People: send opaque `end_user_id` values only. Names, emails and phone numbers are rejected.
- Data: partners are isolated; conversations are never used to train models; DELETE /v1/conversations/{id} deletes permanently; retention 365 days unless agreed otherwise.
- Not yet available: streaming replies, voice through the API, a self-service dashboard, a sandbox, official SDKs.

## VERA CLI

- Early access. `vera "describe the change"` plans it, shows the exact diff and waits for Apply, Revise or Reject. Changes land on a `vera/...` branch as a draft pull request. VERA never deploys and never reads secrets.
- Works with any frontier model provider or OpenAI-compatible endpoint. Connect apps with `vera connect <app>` through the Model Context Protocol.

## Company

- [About](https://veraneural.com/about) · [Safety and privacy](https://veraneural.com/safety) · [Privacy Policy](https://veraneural.com/privacy) · [Contact](https://veraneural.com/contact)