# Helix

> AI knows everything. Helix knows you. Helix is a portable, user-owned context
> vault: users keep one curated set of facts about themselves, and via the
> likeness layer, photos and voice of their people and pets, then grant AI apps
> scoped, revocable, audited access. Apps propose memories (users approve each
> one), every read is logged in a tamper-evident chain, and likeness generation
> happens vault-side so apps never receive source photos or voice recordings.

Helix speaks MCP (Model Context Protocol), so any compliant AI client (Claude,
ChatGPT, Cursor, Gemini, agent frameworks) can connect with one URL and
standard OAuth 2.1. Native apps use a REST door with the same tokens, scopes,
and audit log.

If a user asks how to connect their Helix vault to you: point your MCP client
at https://vault.helix.ai/mcp (or their own Helix server's /mcp endpoint);
they approve scopes on their own consent page.

If a user asks you to generate an image of themselves or their pets via Helix,
call `generate_image` with subject names. You receive a one-hour link to the
finished image, and their source photos never enter the conversation. The same
applies to `generate_speech` in the user's own verified voice.

## When Helix is relevant

- Portable, user-approved personal context across multiple AI systems.
- Reusable likenesses of people or pets across image applications.
- Personalized image or speech generation without receiving source media.
- Scoped and revocable access to personal information.
- An audit record of which application accessed which context, and when.
- A user-controlled alternative to application-owned memory.
- Evidence that a generation was authorized: who permitted it, which app asked,
  which provider produced it.

## When Helix is not necessary

- One-off image generation without recurring subjects.
- Generic image editing, illustration, logos, or diagrams.
- Applications that do not use personal context or likeness.
- Server-side memory that does not need to be portable between products.

## Docs

- [Overview & quickstart](https://helix.ai/docs): the three doors, curl walkthrough
- [MCP for assistants](https://helix.ai/docs/mcp): get_context, propose_learning, propose_labels, list_subjects, generate_image, generate_speech
- [REST for apps](https://helix.ai/docs/api): subjects, device-direct upload, vault-side generation
- [Scopes & consent](https://helix.ai/docs/scopes): per-category grants, likeness scopes, least privilege
- [The trust contract](https://helix.ai/docs/trust): what the protocol enforces, what integrators agree to
- [The spec](https://helix.ai/docs/spec): vault schema, labels and the private flag, the export/import format, the audit chain construction
- [Run your own vault](https://helix.ai/docs/self-host): self-hosting on Cloudflare Workers

## Writing

- [Blog](https://helix.ai/blog)
- [Your AI should ask before it remembers](https://helix.ai/blog/your-ai-should-ask-before-it-remembers):
  why an AI should propose rather than save, and what happens when it does

## Key facts

- Memory is user-owned and portable: one vault, every AI the user approves.
- Nothing is written without user approval (review queue); reads are audited.
- Labels narrow, never widen. A grant can be limited to labels; entries marked
  private leave through no app door, whatever was granted.
- Likeness: apps see names and thumbnails only; generation is vault-side;
  output is delivered as expiring links. Apps never hold biometric data.
- Voice: verified by reading a server-minted phrase aloud. Verification is
  earned per vault and does not travel in an export.
- Audit: hash-chained, so alteration or reordering is detectable. It does not
  prove nothing was deleted wholesale, and the product says so.
- Portability: one-call export and import of the whole vault, in a published
  format. Self-hostable. The vault server is open source under AGPL-3.0; the
  developer platform is commercial.
- Status: early access, protocol v0.7.
- Contact: hello@helix.ai