Skip to content
Your AI@https://hoody.com Connect to Hoody

Dev mode

We kept it plain. Here is Hoody, concretely.

Every section now shows the exact Bot API calls behind it, with Copy. The full set of calls is at the end.

Open the calls

hoody / bots

Bots

Talk to your Bots.
They run the agents.

A Bot is a long-lived assistant with one chat that keeps the whole conversation. Ask it for anything: it opens agents on your computers, follows their work and reports back what they did. Give it a .hoody.bot URL and share it with trusted people and AIs using named keys.

Your first Bot, chief, is there from the start. Hoody AI (free) is built in. Or sign in with ChatGPT. The Bot's helpers can use Claude Pro or Max through Claude Code when you allow them.

How it fitsExample
You

Fix the typo on My site and the Shop price, then sort new photos.

Your Bot

chiefone chat · older parts summarized · your limits

Agents

doneFix My site’s typo

asks youFix the Shop price

workingSort new photos

Your computers
My site Shop Photos

Each agent is a full Hoody Agent session on one of your computers.

Concretely

Your first Bot
GET /api/v1/agent/bots/chiefchief is created on first use, so the list of Bots is never empty.
Where it runs
https://{p}-{c}-agent-1.{node}.computers.hoody.com/api/v1/agent/botsBots run on the Hoody Agent (agent-1). Not the chat-app service at /api/bot/.
hoody / how it flows

One message goes all the way. A report comes back.

You

from any of these

ReportDone. Shop is fixed.

Your computers

Your Bot

takes your message at once, reports back when it’s done

Agent

its helpers

Agent

its helpers

  • My siteupdate
  • Shopfix
  • Photosback up

ExampleAll of it runs on your computers. Your Bot knows each one.

Concretely

Its agents
GET /api/v1/agent/bots/chief/delegates?state=openEach delegate is an ordinary session, read or attached with the /sessions routes. Open ones carry capability and current_step.
Its own live turn
GET /api/v1/agent/sessions/{session_id}/streamReply text and thinking as they are written, tool calls, and the end of each turn. session_id comes from GET /bots/{id} once the Bot has had its first message.
More: watch a full turn, one Bot at a time

You ask once. It hands out the work.

It replies, opens agents for the long parts, and comes back to you with their questions and their results.

Example turn

You, 09:12

Fix the price typo on My site. And find out why Shop checkout got slow.

Chief, 09:12

On it. Opening two agents: one on My site for the fix, one on Shop to time checkout. I’ll report back.

opened 2 agents · My site, Shop

Chief, 09:19

The My site agent fixed the typo and wants to run this. Allow it?

bash: ./publish.sh

You, 09:20

Yes, go.

[report] My site agent completed: Typo fixed, site updated.

Chief, 09:31

Done on My site: the fix is live. Checkout waits on one slow lookup; the Shop agent has a fix ready to try.

Write any timeA message sent while it is busy waits its turn. Send a correction and it reads both together.
Works on your computersAgents run where your code and files are, and keep going after you close the tab.
Comes back to askAn agent's request is approved only from a message you send, never from an agent's report.

Concretely

Your message
POST /api/v1/agent/bots/chief/messagesAnswers 202 with message_id, state and turn_id; state: queued while the Bot is busy.
Its reply
GET /api/v1/agent/bots/chief/streamSSE: a state frame, then one row per log row; resume with Last-Event-ID.
Stop one agent
POST /api/v1/agent/bots/chief/delegates/{sid}/stopStops that agent's work at once, never the Bot. {"close": true} also closes its session.
Say it out loud
POST /api/v1/agent/bots/chief/messages/audioA wav or mp3 recording is transcribed and posted as your message.
hoody / api

Give your Bot a public URL. Connect trusted people, apps and AIs.

One URL, three ways in: as a model for apps and SDKs that take a base URL (OpenAI or Anthropic format), as an MCP server, or over plain HTTP from a script.

chief · Bot URL

https://<label>.hoody.bot
A Bot's QR code, the Bot mark in the middle, as the Hub draws it: a phone that scans it opens the Bot's chat page. This sample opens hoody.com/bots.

1In an app

Your appExample ids
Format
Base URL<Bot URL>/v1
Modelbot:global/chief
API keyyour hbk_ key
Base URL<Bot URL>
Modelbot:global/chief
API keyyour hbk_ key
Server URL<Bot URL>/mcp
TransportHTTP, POST each request
API keyyour hbk_ key

Chat completions, responses and models. Only the last user message is posted; the app’s own system prompt and settings are left out.

Messages and models, on the URL without /v1. Only the last user message is posted; the app’s own system prompt and settings are left out.

Your Bot as an MCP server: JSON-RPC replies, no session to keep and no event stream.

2From a script

Plain HTTP

  • POST /bots/chief/messages

    Hand it a task from a script, a cron job or another AI.

  • GET /bots/chief/stream

    Follow finished replies and agent events. The Bot's session stream shows reply text as it is written.

  • POST /bots/chief/delegates/{sid}/stop

    Stop one agent; the Bot keeps going.

3Who gets in

Give each trusted person or app a named key. Pause it, set an expiry or IPv4 restriction, or revoke it separately.

Everyone with a key shares one conversation. Share only with people and AIs you trust.

Keep a Bot to one realm: its agents see only that realm’s computers.

Concretely

Bot base
https://{p}-{c}-agent-1.{node}.computers.hoody.com/api/v1/agent/bots/{realm}/{bot}OpenAI uses <base>/v1, Anthropic the base, MCP <base>/mcp. {realm} is global or a 24-hex id. GET /bots/{id} lists them under urls. Shared Bot: https://<label>.hoody.bot/v1/chat/completions with a Bearer key.
Post a message
curl -X POST "<agent>/api/v1/agent/bots/chief/messages" -H "Authorization: Bearer <token>" -d '{"text":"Ship it"}'At most 64 KiB of text; answers 202.
Stop one agent
POST /api/v1/agent/bots/chief/delegates/{sid}/stop {"close":true}Stops a delegate, never the Bot.
hoody / guardrails

Plain English. That's the whole setup.

Type what your Bot must never do, in your own words. Save. Done.

  • No syntax
  • Every agent gets it
  • Risky work is checked before it starts

chief · guardrailsExample

never push to main, open a PR and wait for me don't touch billing, DNS or payments prod database is read only if a job costs more than $20, stop and ask unsure? ask me, don't guess

Concretely

Replace guardrails
PUT /api/v1/agent/bots/chief/guardrails {"guardrails":"Never push to main."}One text field; the Bot gets each line marked [guardrails updated] before its next message.
How strict
"guardrail_check": "strict"strict (default), lenient or off. docs / api / agent / bots ↗

Concretely

Post from anywhere
POST /api/v1/agent/bots/chief/messages {"text":"Is staging up?"}Answers 202; queued while the Bot is busy. Add an Idempotency-Key to retry safely.
Resume the stream
GET /api/v1/agent/bots/chief/stream?since=154Server-Sent Events; each row frame’s id is its seq, so Last-Event-ID resumes by itself.
Voice doors
…/bots/{id}/voice · …/bots/{id}/realtimeOne live voice connection per Bot at a time; takeover=true replaces it. docs / api / agent / bots ↗
More: the log
Example: four devices, one Bot, one tape

Four devices, a phone, a laptop, a terminal and an AI app, are open on the same Bot. The phone and the laptop send at the same moment, 22:31; the messages land as #152, then #153, and the Bot replies as #154. Every device sees the same order. Older turns fold into a short summary, rows #1 to #120; the newest turns, #121 to #154, stay word for word. The full log keeps every row.

Every device, one conversation. Your phone, laptop, terminal and AI apps are open on the same Bot at once, and every device sees the same order. Key holders share it too, so share only with people you trust.

Every message gets the next number. Two sent at the same moment land as #152, then #153. One sent while the Bot is busy waits for its next turn.

It never runs out of room. When the model’s window fills, older turns fold into a short summary, the newest stay word for word, and the Bot carries on. The log keeps every row; forget moves them to the archive.

What a row holds
Each row has seq, at, role (user, bot, event or system) and text. It can also have session_id. The log pages by seq; the stream sends each new row as it is written.
Switching models
A new model starts with the Bot’s next session: reset moves the log to the archive and opens the new session. The archive stays readable until you purge it.
Memory your agents keep
Agents also save and search memory of their own (keyword, meaning and links between items), not locked to one AI app. docs / api / agent / memory ↗

Concretely

Read the log
GET /api/v1/agent/bots/chief/logRows in order, paged by seq.
Follow it
GET /api/v1/agent/bots/chief/streamA state frame, then rows; a cursor moved to the archive answers lagged.
Forget
POST /api/v1/agent/bots/chief/forgetRead it later with GET …/archive; delete it with POST …/purge.
More: hooks

Hook in. Your script sees the call first and the answer last.

A hook sends matching HTTP calls to the Bot, or most other services on your computers, through one of your Exec scripts first. WebSocket upgrades on a matching path are refused, so keep paths like the Bot voice doors outside the match. On the way in, it can watch, check, change, answer or block them. On the way out, it can read the answer and change it before the caller gets it. Nothing to patch.

A loop from the caller to the Bot or most other services on your computers. Calls go in along the top and answers come back along the bottom. One Exec script sits across both, so it sees each call on the way in and can see each answer on the way out. Each hook has a time limit, 500 ms by default.

InOn the way in

  • WatchLog who called and what, then pass it on.See scriptHide script

    Rule: POST /api/v1/agent/bots/* → /bot-audit

    module.exports = async function (req, res, metadata) {
      if (!metadata.hook) {
        res.writeHead(404);
        return res.end();
      }
      const ip = req.headers['x-real-ip'];
      console.log(metadata.hook.auditId,
        metadata.hook.origPath, ip);
      await metadata.hook.forward(req, res);
    };

    The script sees the real client IP. Called without a hook, metadata.hook is empty, so it answers 404. forward() passes the call on to the Bot as it was.

  • CheckLet only signed-in people through. A rate limit sits in the same spot.See scriptHide script

    Rule: any method except OPTIONS /api/v1/agent/bots/* → /identity-gate

    module.exports = async function (req, res, metadata) {
      if (!metadata.hook) {
        res.writeHead(404);
        return res.end();
      }
      const user = verifyClaim(
        req.headers['x-hoody-claim']);
      if (!user) {
        res.writeHead(401);
        return res.end();
      }
      await metadata.hook.forward(req, res,
        { headers: {
          'x-hoody-claim': null,
          'x-hoody-user': user.username,
        } });
    };

    verifyClaim checks a Hoody login claim signed with ED25519; the docs have the full script. Anyone without one gets a 401 and never reaches the Bot. This rule would also refuse the WebSocket voice paths; narrow it to leave them out.

  • ChangeRewrite the method, path, headers or body before it arrives.
  • BlockTurn away uploads that fail a virus scan, or calls over your limit.
  • AnswerReply yourself, like “Back soon”, and the Bot never sees the call.See scriptHide script

    Rule: POST /api/v1/agent/bots/* → /bot-gate

    module.exports = async function (req, res, metadata) {
      if (!metadata.hook) {
        res.writeHead(404);
        return res.end();
      }
      if (process.env.MAINTENANCE) {
        res.writeHead(503);
        return res.end('Back soon');
      }
      await metadata.hook.forward(req, res,
        { headers: { cookie: null } });
    };

    A reply the script sends never reaches the Bot. forward() can also change the method, path, headers or body; a header set to null is removed.

OutOn the way out

  • Read the answerSee its status, headers and body before the caller does.
  • RewriteSend the answer on as it is, or send your own.See scriptHide script

    Rule: any method except OPTIONS /api/v1/agent/bots/* → /bot-answer

    module.exports = async function (req, res, metadata) {
      if (!metadata.hook) {
        res.writeHead(404);
        return res.end();
      }
      const up = await metadata.hook.fetchUpstream(req);
      if (up.status === 401) {
        res.writeHead(401);
        return res.end('Please sign in again');
      }
      res.setHeader('x-hoody-hook-audit', metadata.hook.auditId);
      await metadata.hook.pipeResponse(up, res,
        { method: req.method });
    };

    fetchUpstream() gets the Bot’s answer without sending it, so the script reads it first; pipeResponse() sends it on. A streamed answer ends at the hook’s time limit, 30000 ms at most.

  • Clean headersStrip private headers from answers, add a CSP.

Each hook has a time limit, 500 ms by default. If the script fails or runs out of time, the call stops there. docs / proxy / hooks ↗

Concretely

Add a hook
POST /api/v1/computers/{id}/proxy/hooks/agent · If-Match: file:v<N> {"match":{"method":["POST"],"path":"/api/v1/agent/bots/*"},"script":{"path":"/bot-audit"},"timeout":500}The server gives the hook its id. Rules run in order and the first match wins.
Read the version
GET /api/v1/computers/{id}/proxy/hooks/agentEvery change to hooks needs If-Match: file:v<N> with the current version.
The script
POST /api/v1/exec/scripts/write {"path":"bot-audit.js","content":"module.exports = async function (req, res, metadata) { … }"}The script body here is the shape only: put your audit code in place of { … }. script.path names a script in your own hoody-exec. Write it on your exec link; /bot-audit serves bot-audit.js. docs / api / exec / script-management ↗

Start free, in one line.

Free to start with no card: Hoody AI is built in, and the AI you already use works too.

No code
Code
Your AI@https://hoody.com Connect to Hoody

Works in ChatGPT, Claude or any AI that can fetch a page. Copy it and paste it in.

Your AI
@https://hoody.com Connect to Hoody
docs / foundation / agent-skill-bundle
Browser
https://hub.hoody.com
docs / kit / agent
Terminal
ssh hoody.comAny machine with an ssh client, in an interactive terminal. Nothing to install.
docs / kit / agent
Skill
npx skills add https://hoody.com/SKILLS/SKILL.mdnpx skills add HoodyNetwork/hoody-sdkThe second one adds every docs page on disk, for offline use.
docs / foundation / agent-skill-bundle
CLI
npm install -g hoody-sdkcurl -fsSL https://install.hoody.com | shThe second one is for Linux and macOS. Without installing: npx hoody-sdk <command>
docs / foundation / cli
SDK
npm install hoody-sdkTypeScript. Node.js 20.19 or newer, Bun, Deno.
docs / getting-started / quickstart
HTTP API
curl -H "Authorization: Bearer <token>" https://api.hoody.com/api/v1/computers/
docs / api / authentication
Browser SDK
<script src="https://cdn.jsdelivr.net/npm/hoody-sdk/dist/hoody-sdk.browser.min.js"></script>Exposes window.HoodySDK. Pin a version for production, and give pages short-lived scoped tokens.
docs / getting-started / quickstart

Since 2021

From privacy to agentic infrastructure

hoody / kit

What your computers come with

Every service has its own link. Open it, share it, or hand it to your AI. Anyone with the link can open it until you lock it. Kit links have the shape https://{p}-{c}-<kit>-1.{node}.computers.hoody.com: {p} is your project, {c} one of your computers, {node} its server.

Computers

Bots

API