Raynard
A desktop AI agent that builds the tools it needs.
Raynard is a chat app for questions that need real data. Ask it something, and
it answers from services it can already reach. Ask it something it has no way to
reach, and instead of guessing it tells you what connection is missing and
offers to write one — you approve, it codes and tests the extension, and then it
answers the original question.
Runs on macOS, Linux, and Windows. Bring your own model provider.
What it does
Answers from live data, with its sources. Replies carry citations and
structured result cards, so a claim can be checked rather than taken on trust.
Writes its own extensions. When a capability is missing, Raynard asks
permission to build it. A separate coding agent then writes a TypeScript
extension — client, tests, README, and result cards — inside its own directory.
Nothing ships until it passes mocked tests and one real call against the live
API, because mocked tests alone can be green while every endpoint is wrong.
Two visible modes. Ordinary questions run in Explore. Only the
build confirmation moves the app into Build, and the next message returns to
Explore. You always know which one you are in, and extension changes never
happen behind your back.
Steerable mid-answer. You do not have to wait for a turn to finish. Enter
queues a correction the agent picks up at its next tool-call boundary;
Alt+Enter queues a follow-up for when it would otherwise stop. Queued text
sits above the composer until the agent actually takes it.
Work on a schedule. Recurring Explore tasks run daily, weekdays, weekly,
monthly, quarterly, or yearly in your time zone, writing results into a
dedicated chat or an existing one. Scheduled runs can never enter Build on
their own.
Shareable answers, with no server. Sharing an answer builds a link that
carries the whole payload in the URL fragment — which browsers never transmit.
Nothing is uploaded, nothing is stored, and nothing identifies the sender.
Concurrent chats. Several conversations can be working at once. Navigating
away does not cancel a run, and coming back reconnects to it live.
Also: bookmarks, a bundled catalog of pre-approved extensions, token and quota
reporting, and in-app updates.
Private and internal APIs
Nothing about this is limited to public APIs. Point Raynard at your company's
internal service or anything behind a key, give it the documentation — a URL, or
the spec pasted into the chat — and it writes a client for that instead.
Authentication is handled without the secret ever reaching the code. The
extension declares which credentials it needs in its manifest and reads them by
name; the first time a tool runs without one, the app stops and asks you for it
by name rather than failing. The value goes into your operating system keychain,
never into the extension's files, so an extension can be read, edited, or shared
without carrying a key with it.
Each extension runs in its own process and is handed only the credentials it
declared — one extension cannot read another's. Credential values are stripped
out of error messages and failed HTTP responses before they reach the
conversation.
Extensions you build this way stay on your machine. They live in your app
data directory, are never uploaded anywhere, and are not candidates for this
repository's bundled catalog unless you explicitly decide to propose one.
One thing to be clear about: the responses from an API you connect do go to
your model provider, because reading them is how the agent answers your
question. The credentials stay local; the data they unlock does not. Connect
accordingly.
Install
Download the latest release for your platform:
Or install from a terminal, with checksum verification:
# Linux
curl -fsSL https://github.com/geirfreysson/raynard/releases/latest/download/install.sh | sh
# Windows
irm https://github.com/geirfreysson/raynard/releases/latest/download/install.ps1 | iex
You do not need Node.js, Rust, or Homebrew — the packaged app embeds its own
Node runtime. macOS builds are signed and notarized and require macOS 13 or
later. Windows installers are not yet signed, so SmartScreen will warn on first
run.
Connect a model
On first launch Raynard asks for a provider before anything else. You can sign
in with ChatGPT, or use an API key from Claude or Kimi. One provider serves both
the chat and coding roles.
Type /models at any time to switch. API keys are stored in the operating
system keychain — the app's own config file holds preferences only, never
secrets.
Slash commands
| |
|---|
/models | Connect or switch model providers |
/extensions | Browse installed and bundled extensions |
/settings | App version and updates |
/status | Token usage and provider quota |
Development
Requires Node.js and a Rust toolchain.
npm install
npm run tauri dev
Before committing:
npm test -- --run
npm run build
cargo test --manifest-path src-tauri/Cargo.toml --lib
cargo check --manifest-path src-tauri/Cargo.toml
The documentation site in docs/ is a separate Docusaurus project:
cd docs && npm install && npm start
여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗
