How much does Caffeine cost?

Free gives 15 welcome credits for remixing App Market apps. Host is $4 a month billed yearly ($48) or $5 monthly with 20 credits a month; Studio is $20 a month yearly ($240) or $25 monthly with 100 credits; Business is $80 a month yearly or $100 monthly for two seats and 400 credits, or $250 monthly for five seats and 1,000 credits, with extra seats at $50. A credit costs $0.50, top ups run from 100 credits for $50, and going live costs 4 credits.

Does Caffeine support MCP and tool calling?

Full under the Agentic Index verification standard. Caffeine publishes public evidence of Model Context Protocol (MCP) support, tool calling, and API or SDK extensibility for connecting agents to external tools. Evidence basis: The Caffeine MCP server, labeled alpha and free to connect, runs over streamable HTTP with OAuth sign in and sessions that last 90 days, and its tools list, create, preview, publish and delete projects, switch workspaces, report credit usag. Graded from public evidence only against the Agentic Index 14 point capability taxonomy. No vendor pays for placement.

Does Caffeine require human approval before an agent acts?

Full under the Agentic Index verification standard. Caffeine publishes public evidence of human approval steps and guardrails, so an agent action can be held for review before it executes. Evidence basis: Every project has separate draft and live addresses and programs, and going live is a manual step. Version history reverts drafts to earlier versions, a migration check refuses any update that would lose live data, and the agents ask before. Graded from public evidence only against the Agentic Index 14 point capability taxonomy. No vendor pays for placement.

Can you bring your own model or route between providers with Caffeine?

Full under the Agentic Index verification standard. Caffeine publishes public evidence of bring your own model and bring your own key (BYOK), and of routing tasks between models or providers rather than being locked to one. Evidence basis: A model picker sets the model for every agent in a project: Default, Claude Opus 5.5, Claude Fable 5.1, Claude Sonnet 5, GPT 6 Astra, GPT 6 Sol, Kimi K3 or DeepSeek V4.1 Flash, each priced per million tokens in credits. Built apps can also. Graded from public evidence only against the Agentic Index 14 point capability taxonomy. No vendor pays for placement.

Back to vendors

Caffeine

Q: Who is Caffeine for?

founders building a first app, teams building internal tools, educators and students, buyers who want Swiss ownership and chosen hosting machines

Also known as: caffeine.ai, Caffeine AI, Caffeine App Market, Caffeine MCP server

Visit site

Coding agentindependentVerified 2026-11-08

Entry priceFree with 15 welcome credits for remixing App Market apps, or Host at $4 a month billed yearly ($48), or $5 month to month, with 20 credits a month and one live app.Full pricing detail

Caffeine is an app builder from Caffeine Labs AG of Zurich, built by the team behind the Internet Computer, where a team of AI agents turns a chat into a full stack web app and runs it on the Internet Computer. A Quality agent and a browser tester check each build before a draft goes up, going live is a manual step, and customers pick the model from Claude, GPT, Kimi and DeepSeek options. Plans run from free to $250 a month.

Caffeine is an AI app builder sold by Caffeine Labs AG, a Swiss company in Zurich built by the team behind the Internet Computer blockchain at DFINITY, with Dominic Williams named as founder. A customer describes an app, pastes a URL, uploads a zip or imports from GitHub, and Caffeine writes a spec, builds a React and TypeScript frontend with a Motoko backend that stores data inside the program, and deploys it to a live address on the Internet Computer. It is listed on AWS Marketplace, and an App Market carries hundreds of community apps to remix.

A Composer agent runs the build in waves, handing work to Discovery, Product, Design, Frontend, Backend and Quality agents, with frontend work in parallel, compile checks between waves and budget modes that conserve, finish or stop hard. The Quality agent grades each requirement as pass or fail, and a Playwright based tester opens the app in a browser to catch blank screens, broken layouts and console errors before the draft deploys.

Customers pick the model for every agent in a project, from Claude Opus 5.5, Fable 5.1 and Sonnet 5 to GPT 6 Astra and Sol, Kimi K3 and DeepSeek V4.1 Flash. Built apps call outside services over HTTPS, take Stripe payments, send email, post to X and reach users' Gmail and Calendar, and assistants such as Claude can drive Caffeine through its MCP server, which is labeled alpha. Drafts and live apps run separately, Go Live is manual, and version history reverts drafts, though a live deployment has no rollback.

Free gives 15 welcome credits for remixing App Market apps. Host is $4 a month billed yearly, or $5 month to month, with 20 credits and one app; Studio is $20 a month yearly or $25 monthly with 100 credits and unlimited apps; and Business starts at $80 a month yearly for two seats and 400 credits, or $250 a month for five seats and 1,000 credits, with roles, admin controls and OIDC single sign on.

A credit costs $0.50, going live costs four credits, and running an app draws credits for upkeep. Platform sign in goes through Internet Identity, Caffeine names no SOC 2 or ISO attestation, apps cannot be pinned to a subnet or country by default, and accounts with a second deployment cloud or a personal cloud engine choose where apps run.

It fits founders, educators and teams building internal tools who want a running app without setting up hosting, and buyers drawn to Swiss ownership and machines chosen by country. It fits less well for teams that need common SaaS connectors today, since Salesforce, HubSpot, Jira and Slack are listed as coming, for anyone who needs a rollback on a live app, or for buyers who want a named security attestation.

Vendor details

Canonical URL

https://caffeine.ai

Subcategory

Multiagent app builder that writes full stack web apps from chat and runs them on the Internet Computer

Funding status

Private Swiss company, Caffeine Labs AG of Zurich, built by the team behind DFINITY's Internet Computer, with Dominic Williams named as founder. Caffeine gives no funding figures.

Company status

independent

Use cases & customers

Primary use cases

building full stack web apps from a chatinternal tools with stored data and sign inremixing community apps from the App Marketapps with Stripe payments and emaildriving app builds from Claude through MCP

Target customers

founders building a first appteams building internal toolseducators and studentsbuyers who want Swiss ownership and chosen hosting machines

Deployment options

Hosted on the Internet ComputerPersonal cloud engines or a second deployment cloud where the account has oneCode export as a zip or to GitHubCustom domains on paid plans

Integrations

HTTPS outcalls with GET and POST from app backends, Stripe payments, email through Amazon SES, posting to X with OAuth, OpenAI and Caffeine Inference, and each user's Gmail and Google Calendar through their own sign in. Slack posting is experimental, and Google Drive, Microsoft Teams, Outlook, Salesforce, HubSpot, Jira, Confluence, Notion, Dropbox and QuickBooks are listed as coming. A Caffeine MCP server and a command line tool let assistants and scripts drive builds, and DFINITY's ICP MCP connector lets assistants use built apps.

In practice

An operations lead describes an equipment checkout tool in Guided mode, answers three rounds of questions, and watches the agents build it in waves. The Quality agent flags a missing return date field, the agents fix it, and the lead reviews the draft before clicking Go Live.

A teacher remixes a quiz app from the App Market, changes the questions and branding in chat, and publishes it on the school's own domain on the Host plan.

A founder working in Claude asks it to start a Caffeine project for a booking page with Stripe checkout, follows the build transcript through the MCP server, and publishes the result once the draft passes its checks.

Sources & related URLs

Related / legacy domains

https://help.caffeine.ai https://caffeine.ai/pricing https://blog.caffeine.ai/posts/v3-multi-agent https://caffeine.ai/hosting

Research sources

Agentic Index coverage score

11.0 / 14 capabilities · 79%

Integrations & Tool Calling	Full
App backends call outside services over HTTPS with GET and POST requests and custom headers on every plan, and ready skills add Stripe payments, email sent through Amazon SES, posting to X with OAuth, OpenAI and Caffeine's own inference, with each user's Gmail and Google Calendar reachable through that user's own sign in. Slack posting is labeled experimental, and Google Drive, Microsoft Teams, Outlook, Salesforce, HubSpot, Jira, Confluence, Notion, Dropbox and QuickBooks are listed as coming. Built apps cannot hold WebSockets, use Twilio or connect to outside databases or sign in providers. Sourcehelp.caffeine.ai/hc/en-us/articles/46899784385940read 2026-11-08
Workflow Orchestration	Full
A Composer orchestrator runs each build in waves, passing work to specialist Discovery, Product, Design, Frontend, Backend and Quality agents, with frontend work split across agents in parallel and compile checks between waves. Guided mode, the default, asks up to three rounds of questions before building while Instant mode starts at once, budget modes choose whether a build conserves credits, finishes the job or stops hard at a limit, and the agents search the web while they build. Sourcehelp.caffeine.ai/hc/en-us/articles/46899783755924read 2026-11-08
Knowledge Grounding & RAG	Partial
The Data Intelligence agent queries a live app's stored data during a chat, and Studio and Business plans can ask questions across all of a customer's apps. Chat attachments of up to 15 MB each stay with the project, with non image files kept across the conversation and images and PDFs only while recent, and every build rescans the project from scratch. There is no maintained index over the customer's documents for the agents to search. Sourcehelp.caffeine.ai/hc/en-us/articles/46899783755924read 2026-11-08
Human Oversight & Guardrails	Full
Every project has separate draft and live addresses and programs, and going live is a manual step. Version history reverts drafts to earlier versions, a migration check refuses any update that would lose live data, and the agents ask before rerunning the build when the Quality agent reports issues. A live deployment has no rollback and cannot return to a version older than the live one, and through MCP an unanswered question lets the agent decide on its own after five minutes. Comments, password protection and taking an app offline round out the controls. Sourcehelp.caffeine.ai/hc/en-us/articles/46899797134612read 2026-11-08
Security, Identity & Governance	Partial
People sign in with Internet Identity through Google, Microsoft or a passkey, and Business workspaces add OIDC single sign on through Okta, Azure AD, Google Workspace or Auth0, without a way to enforce it, plus Commenter and Builder roles for collaborators and admin controls for up to 100 members. Data is encrypted in transit and at rest, and anonymized prompt content may train Caffeine's AI. Caffeine names no SOC 2 or ISO attestation, and links to uploaded files are not individually authenticated. Sourcehelp.caffeine.ai/hc/en-us/articles/51459220261396read 2026-11-08
Observability & Auditability	Partial
A live checklist and a scrolling reasoning summary show each build as it runs, transcripts of build sessions come back through the MCP server, and credit usage breaks down by charge type and project, with team usage visible to admins. Built apps get visitor, session, pageview, duration and country analytics over 7, 30 or 90 days on Host and above, and Business adds insights across all of a team's apps. Sourcehelp.caffeine.ai/hc/en-us/articles/46899821857044read 2026-11-08
Memory & State Persistence	Partial
A lightweight file per project keeps the user's preferences, build commands that worked and lessons the agents gathered, alongside a spec file and a design brief, and the agents also pick up facts about the person across projects, though project memory is not shared between projects. That file comes with no stated lifetime or delete path. App data persists inside each running program across upgrades, which is how built apps keep their records rather than agent memory. Sourcehelp.caffeine.ai/hc/en-us/articles/46899783755924read 2026-11-08
Deployment & Data Residency	Partial
Apps run as a frontend and a backend program on the Internet Computer with no horizontal scaling, and by default an app cannot be pinned to a subnet or a country. Accounts with more than one deployment cloud choose one in project settings, locked while the app is live, and personal cloud engines join through a registration link with going live at 0.5 credits instead of 4, while OpenCloud machines chosen by country and operator are in early access. Code leaves as a zip or a GitHub export, and self hosting needs Internet Computer tooling. Sourcehelp.caffeine.ai/hc/en-us/articles/46899797134612read 2026-11-08
Prebuilt Agents / Templates / Packs	Full
The App Market holds community apps to browse and remix on every plan, hundreds of them in categories such as education and productivity, and Studio customers can publish their own there. Official templates cover an online store, a cafe and restaurant website and a chatbot, and ready skills add Stripe, object storage, role based access, invite links and RSVPs, user approval and camera and QR features to a build. Sourcecaffeine.ai/app-marketread 2026-11-08
Triggers & Channel Coverage	Partial
Building starts from the web chat, from the Caffeine MCP server inside an assistant such as Claude, or from the command line tool, with a description, a pasted URL, a zip upload or a GitHub import as the starting point. No public API, webhook or schedule starts the builder on its own, so every build begins with a person. Apps it builds can poll an outside API on a timer, as Caffeine's own prompt examples show, but they take no inbound webhooks. Sourcehelp.caffeine.ai/hc/en-us/articles/49663075113236read 2026-11-08
Model Flexibility & Routing	Full
A model picker sets the model for every agent in a project: Default, Claude Opus 5.5, Claude Fable 5.1, Claude Sonnet 5, GPT 6 Astra, GPT 6 Sol, Kimi K3 or DeepSeek V4.1 Flash, each priced per million tokens in credits. Built apps can also call Caffeine Inference, an OpenAI compatible API in early access that routes requests without letting callers pin a model, at $0.20 per million input tokens and $0.50 per million output. Sourcehelp.caffeine.ai/hc/en-us/articles/46899783755924read 2026-11-08
APIs / SDKs / MCP Extensibility	Full
The Caffeine MCP server, labeled alpha and free to connect, runs over streamable HTTP with OAuth sign in and sessions that last 90 days, and its tools list, create, preview, publish and delete projects, switch workspaces, report credit usage and start, steer and read build chats. It is listed in Claude's connector directory, and a command line tool on npm drives the same work, with a local MCP mode in early developer preview. Caffeine has no public REST API or SDK. Sourcehelp.caffeine.ai/hc/en-us/articles/49663075113236read 2026-11-08
Testing, Debugging & Optimization	Full
The Quality agent grades each requirement in the spec as pass or fail, compile checks run between build waves, and when the Quality agent reports gaps the agents ask before rerunning the pipeline to close them. A data loss check refuses any backend update that would drop live data and the agents retry it another way, so every draft clears these gates before it deploys and every change to a live app clears the migration check. Sourcehelp.caffeine.ai/hc/en-us/articles/46899783755924read 2026-11-08
Browser / Computer-use	Full
A Playwright based tester opens a local copy of each app in a real browser, walks through its main flows and checks for blank screens, broken layouts, missing design tokens and console errors, capturing screenshots and console logs for the Quality agent before the draft deploys. Visual testing ships as one of the built in skills, and the customer watches the app in a side by side preview while it builds. Sourcehelp.caffeine.ai/hc/en-us/articles/48917744833300read 2026-11-08

The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded

Pricing

Free with 15 welcome credits for remixing App Market apps, or Host at $4 a month billed yearly ($48), or $5 month to month, with 20 credits a month and one live app.

Monthly or yearly plans that include a monthly credit allowance, with credits at $0.50 each spent on building, going live, app upkeep, email and images.

Free tier

Included quota

Host includes 20 credits a month, Studio 100, Business 400 for two seats or 1,000 for five seats. Balances cap at 22, 110 and 1,100 credits.

What is public

Every plan's yearly price and credit allowance, credit and top up prices, balance caps, and the credit cost of going live, upkeep, email and images.

Billing mechanics

A plan billed monthly or yearly that refills a credit balance each month, with credits spent per build token, per go live, per unit of compute upkeep and per email or image.

Cost watchouts

Running apps draw credits for upkeep every month, going live costs 4 credits each time, larger models burn credits faster, and the free plan cannot go live or build from scratch.

Variable cost rationale

Plans include a fixed credit allowance, and credits pay for building by tokens at the chosen model's rate, for going live, for keeping apps running and for email and images. Larger models, more live apps and heavier app traffic draw credits faster, and top ups refill at $0.50 a credit up to a balance cap.

Additional watchouts

Monthly prices sit in page data rather than in the visible table, and Business appears at two seat and five seat sizes.

Overage / add-ons

Top ups cost $50 for 100 credits, $100 for 200 and $250 for 500. Going live costs 4 credits, or 0.5 on a personal cloud engine, app upkeep draws 0.5 credits per 100 billion cycles, email 0.01 credits per recipient and an image 1 credit.

Sales call required

No, self serve available

Free / trial

Free gives 15 welcome credits and three projects for remixing App Market apps, without going live or building from scratch. There is no free trial, and new Studio subscribers pay $10 for the first month.

Lowest paid plan

Host at $4 a month billed yearly, or $5 month to month

Commercial notes

Caffeine Labs AG of Zurich sells Caffeine, built by the team behind DFINITY's Internet Computer, and lists it on AWS Marketplace.

Key ambiguities

The visible pricing page shows only yearly prices and a two seat Business plan, while the Help Center prices Business at $250 a month for five seats, and an older pricing FAQ still describes three plans and a June launch.

Cancellation / refund

Monthly or yearly plans, and the terms make paid subscriptions non refundable.

Support SLA / resale

Business adds a shared team workspace, roles and permissions, admin controls, OIDC single sign on and a dedicated Caffeine specialist.

Missing data

How many credits a typical build uses for each model.

Agentic Index verified 2026-11-08

Official pricing page Docs

Data confidence: medium