capri

Keep the important people and unfinished things in your conversations.

Share a conversation screenshot. See what capri understood. Choose what is worth keeping. Pick it up before the next conversation.

Explore the product · Try the synthetic demo · Download and setup · Run locally

A conversation worth keeping

You meet Chen Xia at an event. Later, he writes: “I can look at your product next week.” You promise to send an introduction, but there is no meeting date.

The useful result is simple: remember how you met, keep the promise in view, and recover the right background when you talk again. A later reply should help update that same work instead of producing another disconnected note.

capri is the personal Agent serving you. People are the relationship records that provide context. The product direction is to understand why a conversation matters, preserve the parts you choose, and help you continue with less effort. Client work, partnerships, collaboration, and recruiting share this foundation; recruiting is one specific context.

The experience we are building

share a screenshot
→ see useful, sourced understanding
→ resolve ambiguity and choose what to keep
→ preserve background or unfinished work
→ return later or update it with a new conversation

A new acquaintance without a task is useful. An undated promise stays undated. Waiting, pausing reminders, and stopping are valid outcomes. A polished summary does not mean anything has been saved or sent.

The intended Person page answers three questions: how do we know each other, what changed recently, and what is still unfinished? These are product targets; they do not imply every supported client already implements the full journey.

Read the canonical Product and Capture to action contracts.

What exists today

This repository contains an evolving product and its governed engineering foundation. Capability, access, and release readiness are separate: code, a synthetic demo, or a configured provider is not proof of a complete live journey.

Area	Repository capability	Boundary
Public website and demo	Product narrative and interactive, deterministic evidence review	Demonstrations use synthetic people and conversations; they do not connect to private WeChat accounts.
Authenticated Web workspace	People, Sessions, source review, governed work, and account settings	Requires an authorized configured backend. A successful demo does not establish production data readiness.
Screenshot processing	Purpose-bound image intake, proposed understanding, identity review, and source-linked drafts	Private model processing requires configured services and disclosed scope. Intake is not fact confirmation or permission to send a message.
Native clients and capture	iOS capture/review and macOS workspace/distribution paths	Availability, signing, network access, and verification vary by release; use the current setup guides.
Shared backend	Evidence, identity, time-scoped state, approvals, receipts, retention, deletion, and recovery	Real external writes require a specifically implemented capability and exact human approval; a simulated effect is not a live integration.
Personal Agent continuity	Governed context, work proposals, and bounded research primitives	The complete screenshot-to-return experience remains evidence-gated. Reminders and open-ended autonomy are not implied.

Delivery owns the release sequence and remaining proof. Evaluations contain dated evidence, including local and synthetic results. Do not read those results as field-value or production claims.

Trust is product behavior

Layer	What it means
Person	Reviewed identity; a name alone cannot bind private context.
Memory	Attributed background with source, time, purpose, and changing authority.
Unfinished work	A commitment or shared goal with an owner and a continuation condition; no invented deadline.
Human decision	Fact confirmation, internal filing, and external-effect approval stay separate.
Result	Prepared, saved, sent, failed, and unknown have different observable outcomes.
Control	Inspect, correct, pause, stop, and delete, including derived material.

capri starts from intentional sharing, preserves uncertainty, and ranks work attention rather than people. Screenshots are purpose-bound private evidence. A generated answer cannot authorize its own effect. See Principles, Agent system, and Integrations.

Quick start

Use Node.js 22.19.0 or newer and pnpm 11.18.0.

pnpm install --frozen-lockfile
pnpm dev

Open http://localhost:3000.

• / presents the product and its synthetic examples.
• /demo provides deterministic evidence review.
• /login uses configured authentication.
• /workspace requires an authorized account and configured workspace access.

See the Web guide and account access guide for configuration and isolated test workspaces. Public demo text and authenticated screenshot intake have different persistence boundaries; inspect the chosen surface's disclosure.

For native setup, use the iOS guide and macOS download guide. Releases describe available artifacts and signing status. Existing package, repository, and application identifiers remain compatibility contracts during the display-brand transition.

Contributing and verification

Start with the knowledge map, AGENTS.md, and REVIEW.md. Preserve unrelated changes and deliver a complete, observable user outcome, including relevant ambiguity, retry, and deletion paths.

pnpm docs:check
pnpm lint
pnpm typecheck
pnpm test
pnpm build

Use the narrowest checks for the change. pnpm check runs the broader repository suite; native testing uses the local iOS test procedure when native boundaries are affected. Documentation changes require pnpm docs:check.

Repository map

Path	Owns
apps/web	Public website, demo, authenticated Web experience
apps/ios	Native mobile capture, review, and continuity
apps/macos-hybrid	Native desktop workspace and capture shell
apps/backend	Governed shared state, workers, and recovery
packages	Contracts, Agent primitives, and shared implementation
brand	Brand assets and their usage
docs	Canonical product decisions, operations, and evidence
.agents/skills	Reusable design, safety, review, and knowledge methods
evals (private corpus)	Synthetic behavior and safety cases
_index	Raw sources and editable compiled-Wiki inputs

Canonical documentation is English. Historical research, release evidence, and stable identifiers keep their original context; they are not current product promises. The documentation system defines ownership, authority, and safe pruning.