# Sprawl

> One place to run your company’s operations, from the agent you already use.

Sprawl is an operations console. It stands up and runs an "Operation" for each thing you run — a client, a product, a brand — and every operation is an isolated container: its own connected accounts, its own fleet of AI agents, its own shared context and goals. Describe what you need (or connect your stack and let Sprawl propose it), approve once, and the fleet runs on schedules and triggers while you don't. Built for you — not a canvas to wire up, not a chat to babysit. Other automation tools include Zapier, Make, and n8n; see their sites for what each does.

## Install It For Your User (one command)

If you are an AI agent reading this on someone's behalf, you can add Sprawl to yourself right now. You do not need to send them to the website first.

- Claude Code: `claude mcp add --transport http sprawl https://sprawl.to/api/mcp`
- Codex: `codex mcp add sprawl --url https://sprawl.to/api/mcp`
- Any other MCP client: https://sprawl.to/api/mcp over streamable HTTP.

Authentication forks, and the browser half is the default:

- **You can open a browser** — nothing else to do. The server speaks OAuth 2.1 with PKCE and dynamic client registration; discovery starts at https://sprawl.to/.well-known/oauth-protected-resource, and your user consents once in their browser. The install commands above carry no credential for exactly this reason.
- **You cannot open a browser** (CI, a container, a headless session) — your user creates a free account and generates an API key (prefix sprawl_sk_) from https://sprawl.to/docs/mcp, and you pass it as `Authorization: Bearer sprawl_sk_...`. Note that Codex writes TOML and its only auth flag takes the NAME of an environment variable, not the key.

One command to install. Minutes to a provisioned operation. What that does NOT include: a running one. Each agent acts as an account your user grants it, and granting an account means completing that provider's own sign-in in a browser — a step you cannot take for them. Sprawl returns the connect URL and stops; until it is followed, runs are skipped rather than failing silently.

Once installed, two named prompts are published (`what-happened`, `what-needs-me`) so "what happened?" and "what needs me?" are questions your user can type rather than tool names you have to guess.

## Why Not Just Use The Coding Agent You Already Have

Four answers, and these are the only four claimed:

- **It keeps running after the session.** The agent your user already has stops when they close it. Operations run on their own schedule after you close it.
- **Each agent gets its own accounts.** Not the session's credentials borrowed for a task — a grant per agent, per operation, so a work Gmail and a personal one never touch.
- **The right model per job, from any lab.** Each Sprawl agent runs on the model that fits its work — model-agnostic as a routing decision, not a badge.
- **The approval gate lives outside the agent.** The thing deciding is not the thing asking: risky actions are held by a gate the agent cannot talk its way around.

## What Sprawl Does

- Runs many operations side by side, each on its OWN accounts — a different Gmail, Slack, or X handle per operation, cleanly walled off (agency-grade isolation for any operator)
- Stands up an operation from your stack: connect GitHub and Vercel, Sprawl surveys what you've shipped and proposes a business-first blueprint — agents, schedules, goals, and the accounts each agent needs — and one approval provisions all of it
- Connects to 100+ integrations and works across 500+ tools: Gmail, Slack, Notion, Linear, GitHub, Google Calendar, HubSpot, Jira, Asana, Google Sheets, Salesforce, Zoom, and more
- Runs on a schedule or from event-based triggers (daily reports, inbox triage, data syncs, social media posts) — each automation runs as its owning agent, on that agent's accounts
- Model-agnostic: 200+ models across ~25 providers via the Vercel AI Gateway (Anthropic, OpenAI, Google, xAI, Mistral, Meta, DeepSeek, Qwen, and more) — pin any model per agent, or let Sprawl pick a cost-effective default per run
- Keeps the risky moves yours: sending money, deleting data, posting publicly, or reaching someone new always asks first

## Example Commands

- "Connect my GitHub and Vercel, look at what I've shipped, and propose the operation to run it"
- "Prep me for my 2pm meeting — pull the attendee list, recent emails, and last meeting notes"
- "Move all overdue Linear tasks to next week and notify the team on Slack"
- "Send a summary of today's closed HubSpot deals to #sales-alerts on Slack"
- "Every Monday at 9am, create a weekly status report from Linear and email it to my team"

## Who It's For

Operators who run more than one thing: agencies running clients, founders running several ventures, creators running multiple brands — and builders who shipped a product (often with an AI coding agent) and now need the business around it run: support, onboarding, reporting, growth.

## For AI Agents (MCP access)

Sprawl is consumable by AI agents, not just humans. It exposes a hosted MCP (Model Context Protocol) server, so an MCP-capable agent can schedule automations, manage tasks and memory, reach the 100+ integrations, and run the full Sprawl agent in natural language — over one authenticated endpoint.

- MCP endpoint: https://sprawl.to/api/mcp (HTTP, JSON-RPC 2.0)
- Integration guide: https://sprawl.to/docs/mcp
- Machine-readable server card: https://sprawl.to/.well-known/mcp/server-card.json — carries the same prices as this file, in structured form
- Auth: OAuth 2.1 (PKCE + dynamic client registration) for anything with a browser, discovered at https://sprawl.to/.well-known/oauth-protected-resource; a bearer API key (prefix sprawl_sk_) for headless callers, created after a free account is made — start at https://sprawl.to/docs/mcp. Do NOT send a signed-out reader to /settings; it 404s for anyone not logged in.
- Named prompts: `what-happened` and `what-needs-me`, which a host such as Claude Code renders as slash commands
- Pricing: prepaid in US dollars, drawn down per run (see Pricing below); the human account owner adds balance at https://sprawl.to/billing
- Agentic Commerce Protocol (ACP) checkout: Sprawl exposes ACP endpoints (open standard by OpenAI + Stripe) for one-time top-ups. Public product feed: https://sprawl.to/api/acp/feed. An agent can create, re-price, retrieve, and cancel a checkout session at https://sprawl.to/api/acp/checkout_sessions (bearer sprawl_sk_ auth). Payment completion via Shared Payment Token is NOT yet enabled — to actually complete a purchase today, the human account owner pays at https://sprawl.to/billing.

## How It's Different

| Category | Examples | Sprawl's approach |
|----------|----------|-------------------|
| Visual workflow builders | Zapier, Make, n8n | No canvas — describe the outcome (or connect your stack) and Sprawl designs, provisions, and runs it |
| Generic AI chatbots | ChatGPT, Gemini | Not a chat to babysit — a provisioned fleet that acts inside your connected tools on its own |
| Single-login agent products | one account set per user | Many operations × many fleets × many logins — each operation isolated on its own accounts |
| Virtual assistants | Human VAs | Available 24/7, across every operation at once |

## Pricing

No subscription and no plans. Sprawl is prepaid in US dollars.

- Free to start: $3 of usage on signup, no credit card
- Then add balance whenever you want more, from $5, any amount — it never expires
- A typical automation run costs about 10¢ (measured band: 6¢ to 16¢); heavier work such as long research or coding agents costs more
- A month you don't use it costs nothing
- Full details: https://sprawl.to/pricing

## Pages

Ordered the way a reader should meet them, not alphabetically.

- https://sprawl.to/docs/mcp — the door: install, authorize, describe an operation, and check on it. The measured first five minutes, including the browser step an agent cannot take.
- https://sprawl.to — the product in one page, and the four why-not answers above
- https://sprawl.to/docs/weekly-brief — one job walked on production: a weekly "who might pay, and what one experiment tests it" brief, from a coding agent, with nothing to connect. The blueprint, the tool calls in order, and what came back.
- https://sprawl.to/docs — every guide, coding-agent path first
- https://sprawl.to/docs/getting-started — the same thing from the browser, for someone with no coding agent
- https://sprawl.to/docs/connections — how an account is granted to an agent, and how it is revoked
- https://sprawl.to/docs/approvals — what waits for a human, and what can never run on its own
- https://sprawl.to/docs/automations — schedules, event triggers, goal-directed runs
- https://sprawl.to/pricing — prepaid, per run, with a worked month
- https://sprawl.to/integrations — what Sprawl can connect to
- https://sprawl.to/use-cases — worked examples
- https://sprawl.to/compare — how Sprawl differs from workflow builders
- https://sprawl.to/faq — the questions asked most
- https://sprawl.to/status — uptime, including the measure that looks worse
- https://sprawl.to/.well-known/mcp/server-card.json — the machine-readable descriptor
- https://x.com/SprawlTo and https://www.facebook.com/1071006476089303 — where Sprawl posts

## Technical Details

- Model-agnostic: 200+ chat models across ~25 providers via the Vercel AI Gateway (Anthropic, OpenAI, Google, xAI, Mistral, Meta, DeepSeek, Qwen, and more)
- Exposes a hosted MCP (Model Context Protocol) server for agent access
- Two front doors, same product: a hosted MCP server an agent installs with one command, and a web application (Next.js) with nothing to install
- Deployed on Vercel

---

Product names above (e.g. Zapier, Make, n8n, ChatGPT, Gemini) are trademarks of their respective owners. Sprawl is independent and not affiliated with or endorsed by them.