Skip to content

Beta: these docs are still being updated. The next full update is on 18 November at 8 PM UTC. Docs in beta. Full update 18 Nov, 8 PM UTC.

Hoody.com

Lookup tables for The Vibe Coding Revolution.

Every tool the agent needs, and every tool you need to watch the agent, is already a URL of the computer, so the whole workflow lives in browser tabs.

ServiceWhat it gives the session
Hoody Agent (hoody-agent)An autonomous coding agent, controllable over HTTP
Hoody Code (hoody-code)VS Code running in your browser, with nothing to install
Hoody Terminal (hoody-terminal)Shell access from anywhere, with multiple instances
Hoody Displays (hoody-display)A live preview of GUI applications in real time
Hoody Exec (hoody-exec)Scripts that become API endpoints as soon as they are saved
Hoody SQLite (hoody-sqlite)A database you reach through HTTP calls
SnapshotsA checkpoint at any moment; a small restore finishes in seconds

The only setup before you start is a computer. When the agent installs a package, it runs the command in the computer through its shell tool. Creating a table or checking the preview goes through the same HTTP services you can call yourself, which is what lets one agent drive the whole loop.

TabWhat you see while the agent builds
Hoody Code (code-1)Files appear and change as the agent saves them. In the dashboard example it creates src/App.tsx, src/components/MetricsChart.tsx and api/metrics.ts, and you can open each one to see its imports and function definitions land.
Hoody Terminal (terminal-1)A shell in the same computer, where you can inspect what the agent’s commands (bun install react react-dom tailwindcss, bun create vite . --template react-ts, mkdir -p src/components) produced. The agent runs those commands itself and reports their output in its own chat; it does not type into this terminal.
Hoody Displays (display-1)The app preview, after you start its development server and open its local URL in a browser on display 1: first the Vite scaffold, then the layout, then the charts, then real data flowing in.
Hoody Agent (agent-1)The agent TUI, with Chat, Files and Computers tabs. Keep the editor and graphical preview open in separate browser tabs.

The experience is the same on a laptop, a phone or a tablet: the computer is the source of truth, and every display is a view into it.

The display admits one viewer at a time by default, and a new viewer disconnects the current one. Add these to the display URL before you share it (display sharing):

ParameterEffect
sharing=trueSeveral viewers at once
steal=falseA new viewer joins instead of disconnecting the current one
readonly=trueThis viewer watches but cannot type or click

What a prompt:sync call returns with policy=auto_approve:

What happens in the turnWhat the call does
The main agent asks to confirm a tool call (helper confirmations still require a person)Approved for you, so the agent keeps working without pausing, unless the gate marks a destructive command (gate_cause: static_stakes), which is denied
A confirmation is raised by your tool-call rulesDenied
A human-only gateReturns pending_gate for a person to answer
The agent asks a questionReturns pending_gate for you to answer
The session’s approval policy requires a decision on every actionRefused with 409 approval_policy_active
The turn is still running after 290 secondsReturns pending_gate when a gate waits on a person; otherwise answers 503 with details.turn_id, and the agent keeps working. Watch it in the tabs, or poll GET /api/v1/agent/sessions/SESSION_ID/turns/TURN_ID

The call returns an outcome, not the agent’s reply: a finished turn includes status and turn_id; a turn paused for your answer includes pending_gate and turn_id instead. Read the reply in the agent tab or with hoody agent sessions transcript get. The examples create a new session. In the browser agent, open Sessions (F1), select the session the call created and press Enter to join it. Open Chat (F2) to watch its replies and answer its gates. Answer a pending gate in the agent tab, or with POST /api/v1/agent/sessions/SESSION_ID/confirm or /answer (approvals). The CLI stops waiting for a request after 30 seconds, so the CLI tabs dispatch the turn with hoody agent sessions turns create, then poll turns get for its state and the session for a pending_gate. Every outcome and error is on the session turns API page.

AI-generated code is confident code: it compiles, it runs, and sometimes it is still subtly wrong. A snapshot taken before the change makes that risk acceptable.

RiskMitigation
Agent installs bad packagesRestore snapshot
Agent deletes important filesRestore snapshot
Agent breaks the databaseRestore snapshot
Agent introduces security vulnerabilityRestore snapshot, review code
Agent rewrites working code incorrectlyRestore a snapshot (this restarts a running computer), try a different prompt

A restore puts back that computer’s own storage only: work already sent to a remote service, or files on a mounted share, stay as they are. How long it takes: Restore a snapshot.

AgentWhat it does
generalThe default: writes and modifies code, running tools as its configured permissions allow
planWorks through explore, plan and execute phases and produces a detailed plan before implementing
exploreInvestigates the code and reports; its tools only read, so it changes nothing

Hooks fire on session and tool lifecycle events; each runs a shell command you supply, wired to its event in plain JSON configuration. The hooks API lists every event.

HookWhat it can do
UserPromptSubmitInject “CRITICAL: Never execute rm -rf, DROP TABLE, or git push —force without asking first” into every turn
PreToolUse matching bashFire a notification whenever the agent is about to run a shell command, or block the call outright
StopPing Slack whenever the agent finishes a turn
SessionEndPing Slack when a session closes

The two-hour Twitter clone has become the stock example for this style of development. On Hoody, it looks like this:

Hour 1: The conversation

You: "Build a social media app like Twitter. Users can post short messages,
follow other users, and see a feed. Use React, Tailwind, and SQLite."
Agent: [Creates database schema: users, posts, follows tables]
[Scaffolds React app with Vite]
[Implements authentication with session tokens]
[Builds the feed algorithm]
[Creates the post composer component]
[Wires up the API endpoints via hoody-exec]
You: "The feed should show newest first. Add a character counter
that turns red at 280 characters."
Agent: [Modifies feed query to ORDER BY created_at DESC]
[Adds character counter component with conditional styling]

Hour 2: Polish and ship

You: "Add profile pages with a follow/unfollow button.
Make it look professional -- use a card-based layout."
Agent: [Creates /profile/[username] route]
[Implements follow/unfollow toggle]
[Redesigns layout with card components]
[Adds loading skeletons]
[Writes automated tests]
You: "Perfect. Snapshot this and create a production alias."
Agent: [Creates snapshot: twitter-clone-v1]
[Creates proxy alias: my-twitter.node-us.computers.hoody.com]

One conversation produced a working application with a production URL. The AI wrote the code, configured the database, set up the routing and deployed; you provided the creative direction.