Open source · runs on your computer

Tell Todd what you want.
It does the whole thing.

Todd is a team of AI agents that builds, launches and even pays for what your project needs, start to finish. It only stops to ask before it pays by card or posts in your name.

Get Todd See how it works

YouLaunch a waitlist site for dropin.hockey, with sign-in and a launch post.

Todd3 agents working together Done

Built the website
Put it online
Added sign-in with Google and email
Bought dropin.hockey $10.44 · you tapped Approve
Checked the live site for errors
Posted the launch thread you tapped Approve

Your part: 1 message, 2 tapsNo logins, no copy-paste

Free and open source (MIT)
Runs on your computer or server
Your Claude plan or any AI model

What you can hand off

Whole jobs, not just steps.

Launch a product

Website, domain, sign-in, deploy and the launch post.

You: 2 taps

Ship and fix code

New repo, live deploy, then it finds and fixes bugs on the real site.

You: 0 taps

Buy and set up

Domains, services and checkouts, within the budget you set.

You: 1 tap per card payment

Research and organize

Compares options and fills a spreadsheet or a Notion table, with sources.

You: 0 taps

How it works

Sign in once. Ask once. Done.

Todd does every step. You only show up where your rules say so.

1
Sign in once
Log in to your accounts in Todd's browser. Their command-line tools connect too.
GitHubVercelStripeGoogle+65
2
Say what you want
One sentence, in plain words. Add a budget if you like.
Build a booking site for my rink…
3
Todd handles the rest
It plans, splits the work between agents, builds, deploys, buys and checks its own work.

ToddYou

Planspicks the agents
Ask once
Buildscode and content
—
Deployslive and checked
—
Buysdomain, services
Tap Approve
Launchesposts, listings
Tap Approve
Reportswhat it did, links
Read the summary

Want to change something mid-way? Message it any time, or pause it. When a run is done, send another message and it keeps going.

Spending

Yes, it can buy things. On your terms.

Domains, hosting, subscriptions, checkouts. The rules live in Todd's code, so the AI can't talk its way past them.

Small account purchases: automatic

Under your limit ($15 by default), through an account you're signed in to.

Vercel Domains · rinkbook.app$12.00Auto-approved

Card checkouts: one tap

Todd fills in the checkout. Your card only works on that store's own site, after you approve.

$$10.44 at Porkbundropin.hockey, 1 yearApprove

A budget for every run

Going over it needs your OK. Every charge goes into a ledger you can check.

$22.44 of $50

Example runs

Watch it work.

Every agent gets its own live window. Here's what three runs look like in the dashboard.

Launch a waitlist site for dropin.hockey with sign-in and a launch post You: 2 taps

Site & Domainsandbox · github · vercel · browser

Built the site, build passed
New GitHub repo, code pushed
Deployed to Vercel
Opened it live: no errors
Checkout for dropin.hockey waiting for your tap

Firebase Authsandbox · browser

Created the Firebase project
Turned on Google and email sign-in
Sent the config to Site & Domain
Done, with a summary

Launch Assetssandbox · web · browser

Read recent launch posts in the niche
Wrote a 3-post X thread
Made a social card image
Post the thread exact text shown to you first

Approve spending $10.44 at Porkbun?Register dropin.hockey for 1 year. The card only works on porkbun.com.

DenyApprove

Build rinkbook, put it in a new GitHub repo and deploy it to Vercel You: 0 taps

Build & Shipsandbox · browser

Wrote the app, build passed
gh repo create rinkbook --push
vercel deploy --prod
Opened the live site: 1 error, 1 missing file
Fixed it, pushed, redeployed
Checked again: no errors

Liverinkbook.vercel.appTake control

rinkbookDrop-in ice times this week

North Rink · Sat 6:15 AM4 spots

Lakeside Arena · Sat 9:00 PM11 spots

FAILED REQUEST: 404 /rinks.json
# after the fix
No JavaScript errors or console warnings.

Features

Everything a job needs, built in.

A planner that builds the team

Each goal gets its own agents, each with only the tools its part needs. They work in parallel, and the planner keeps listening to you while they do.

Watch and steer

A live window per agent. Pause it, message it, or stop it.

Sign in once, web and CLI

A real browser that checks its work

Agents open what they built, read errors and fix them. Take control any time.

{{secret:STRIPE_KEY}}↓ only inside the commandsk_live_••••••

Secrets never reach the AI

Keys live in an encrypted vault. Agents use them by name; outputs are scrubbed.

Your model, your server

Runs as Claude Code on your Claude plan, or any model with your own keys.

Plugins and MCP

Drop in a Python file or an MCP server and the planner can hand it out.

Picks up where it left off

Every run is saved as it goes. After a restart, press Resume. After it's done, send another message and the same planner continues.

Under the hood

For the curious.

How each piece works, with the tools agents actually call.

Planner and agentsspawn, wait, message, cancel

The planner writes a checklist and calls spawn_agent once per group of related work, with only the toolsets that group needs. Agents run in the background (at most 4 at once and 12 per run by default). A message from you ends wait_for_agents early so the planner can act right away, and it can't finish while its agents are still running.

spawn_agent(name="Site & Domain", task="Build, deploy, buy the domain",
            toolsets=["sandbox", "vercel", "browser"])
wait_for_agents()
message_agent("a7c…", "Also draft a LinkedIn post")

APIs first, browser lastfind_integrations

Before touching a service, an agent calls find_integrations. It picks a ready toolset or MCP server, then the service's REST API with a vault key, then a connected CLI. The browser comes last, and starting it for a service needs a written reason. Checking a site the agents built needs none.

Sign in once, web and CLIcli_login, cli, gh

When you sign in to GitHub, Vercel, Netlify, Railway, Cloudflare, Stripe or Firebase, Todd starts that CLI's login in the sandbox and approves it in the browser with your session: it fills the device code and clicks Continue. Some sites keep the final Authorize click for people; Todd scrolls to it and asks you, once, at setup. The result goes straight into the vault, and agents run the CLI with cli("vercel", "deploy --prod --yes") or gh(…).

The sandboxshell, files, git

A container with Node, Python, git, gh and deploy CLIs. All agents in a run share its workspace. It can't reach the database, the dashboard or the browser, and it never sees the vault: secrets are injected per command. git_push refuses repos whose config could redirect the push, and scrubs the token from output.

The browserlive view, read text, console

One persistent Chromium, shared by all agents and watched live in the dashboard. Browser tasks take turns while other work continues. Agents copy page text with browser_read_text, move keys a page shows straight into the vault with browser_save_secret, and read JavaScript errors and failed requests with browser_console.

Spendingpolicy, budgets, ledger

Every purchase goes through one function, authorize_spend, which checks the auto-approve limit and the run's remaining budget; anything outside them asks you first. Card payments always need your approval, and card details reach the browser only as masked placeholders, only on the approved merchant's own domains, with screenshots off. Posts and messages from your accounts need approval with the exact text.

Engines and modelsClaude Code or LiteLLM

By default each agent runs as a headless Claude Code session on your Claude plan. Or switch to API keys and pick any LiteLLM model per tier: Anthropic, OpenAI, Gemini, OpenRouter, Groq, DeepSeek, or local through Ollama.

Saving and resumingcheckpoints

LangGraph with a Postgres checkpointer saves every run. A run interrupted by a restart shows Resume; a finished run continues when you message its planner, in the same conversation.

Connectors

About 70 services, signed in once.

Plus any site you add. Tags show the faster route Todd uses before the browser.

Toolset built in CLI signs in with your browser MCP one-click server API with a saved key

GitHub ToolsetCLI Vercel ToolsetCLI Stripe CLIMCP Netlify CLI Cloudflare CLI Railway CLI Firebase CLI Supabase MCP Neon MCP Notion MCP Linear MCP Figma MCP Sentry MCP Porkbun API Namecheap API Shopify API Slack API Resend API X API Reddit API Instagram API App Store Connect API Google Play API OpenAI Platform API Anthropic Console API GmailGoogle DriveGoogle CloudSearch ConsoleAnalytics AWSAzureRenderFly.ioHerokuDigitalOceannpmExpo GoDaddyPayPalGumroadLemon SqueezyAmazonEtsyeBay LinkedInTikTokFacebookThreadsYouTubePinterestDiscordTwitchBluesky Product HuntHacker NewsIndie HackersMediumSubstackDEV CanvaCalendlyOutlookDropboxbeehiivMailchimp Apple DeveloperHugging FaceReplicate

Compare

What sets Todd apart.

Only the differences that matter, as of September 2026.

Capability	Todd	Codex	Claude in Chrome	Claude Code
Takes a goal all the way: code, accounts, deploy, launch	Start to finish	Coding tasks	Tasks in your tabs	Coding tasks
Buys what the goal needs	Within your limits	Not built for it	Declines purchases	Not built for it
A team of agents, planned for each goal	Planner builds it	Parallel tasks, subagents	One assistant	Subagents
Signed in to your accounts, web and CLI	Sign in once	GitHub, local tools	Your Chrome logins	Your local tools
Runs on your server, with any model	Any model, even local	CLI: local, many models	Hosted, Claude	Local, Claude models

YesPartlyNo

Where others shine PR reviews Codex, Claude Code Coding in one repo Codex CLI, Claude Code A quick task in your own browser Claude in Chrome

Sources

Codex: cloud, models, subagents, providers, github.com/openai/codex
Claude in Chrome: permissions, safety
Claude Code: agents, on the web, Chrome
Products change fast. Tell us if something here is out of date.

Quick start

Running in five minutes.

You need Docker. A Claude Pro or Max plan, or API keys for any model.

Terminal

git clone https://github.com/bobby-burns/todd.git
cd todd && cp .env.example .env
docker compose up -d --build
open http://localhost:3000

Then the setup wizard walks you through your model, spending limits and sign-ins.

Todd has no dashboard login yet. Keep it on 127.0.0.1, off the internet.

Waitlist

Hear when Todd ships.

Leave your email and we'll tell you about new releases and early access. No spam.

2 people have joined the waitlist

We only use your email to write to you about Todd.

Give Todd a goal.

Open source, on your computer, with your model.

Get Todd on GitHub

Tell Todd what you want.It does the whole thing.