Internet Computer (ICP) and AI Explained: LLMs and Caffeine

Not financial advice. This article explains how a technology works. It is not investment advice and it makes no price predictions. Crypto tokens are volatile and you can lose money. Do your own research and talk to a licensed adviser before making any financial decision.

The Internet Computer (ICP) has been in AI news a lot lately. In the last few weeks DFINITY linked ICP to Claude and ChatGPT, and its Caffeine app builder added a choice of clouds.

This guide explains what ICP is, how AI runs on it, what Caffeine does, and the real limits and risks. It also compares ICP with Bittensor and NEAR in plain English.

TLDR

What ICP is: a blockchain network from the DFINITY Foundation that runs full apps (code, data and web pages) inside “canisters”. Mainnet launched in May 2021.
On-chain AI: the LLM canister gives apps access to Llama 3.1 8B running on ICP nodes. It is free during the rollout, with limits like 1,000 output tokens.
Caffeine: DFINITY’s chat-to-app builder. AI agents write a Motoko backend and deploy it to ICP.
New in late 2026: an official ICP MCP connector for Claude (October 30) and a Caffeine cloud picker (November 7).
Main limits: only small models fit on-chain, and much of the “AI” still uses outside models.
Versus others: Bittensor rewards AI work in subnets, NEAR focuses on private AI in secure hardware, ICP focuses on hosting apps and agents.

What the Internet Computer (ICP) Is

The Internet Computer is a public blockchain network built by the DFINITY Foundation. Its mainnet went live in May 2021. ICP is also the name of its token.

Most blockchains store small smart contracts and send the rest of an app to normal cloud servers. ICP tries to host the whole app on-chain. That includes the backend logic, the data and even the web pages people load in a browser.

Here are the main parts, in plain words:

Canisters: the apps. Each one bundles WebAssembly (Wasm) code with its own saved data.
Subnets: groups of computers (nodes) that run canisters together. A standard application subnet has 13 nodes, and every node runs each update to agree on the result.
Network Nervous System (NNS): the on-chain governance system. People lock ICP into “neurons” to vote on proposals, such as protocol upgrades or adding subnets.
Internet Identity: ICP’s own sign-in system for apps and connectors.
Chain Fusion: features that let canisters work with other chains like Bitcoin, Ethereum and Solana.

If you are new to this whole area, my AI crypto tokens explainer gives the bigger picture first.

Canisters, Cycles and Who Pays

Canisters pay for their own compute, storage and bandwidth with cycles. The network takes cycles from the canister as it runs. Users of the app do not pay per click, a bit like a normal website.

Item	How it works
Cycle price	1 trillion cycles = 1 XDR (the IMF’s Special Drawing Right), so costs stay steady when the ICP token price moves
Getting cycles	ICP is converted in the Cycles Minting Canister. Cycles cannot be turned back into ICP.
Who pays	The developer tops up the canister. Anyone can add cycles to it.
Running out	The canister freezes and stops answering. If it stays empty, it is eventually deleted with its data.
Memory	Heap up to 4 GiB (wasm32) or 6 GiB (wasm64); stable memory up to 500 GiB

Note: Those memory limits matter a lot for AI. A large language model’s weights must fit inside a canister, which is why only small models run fully on-chain today.

How AI Runs on ICP: Three Approaches

“AI on ICP” can mean three quite different things. It helps to keep them apart.

1. The LLM canister (shared on-chain model)

DFINITY runs a shared system service called the LLM canister. Your canister sends it a prompt and gets an answer back, with no API key. The docs say inference runs on ICP nodes and is seeded from ICP’s random beacon, so results are deterministic and checkable by the subnet.

Here is a Rust example from the official docs. The ic_llm library also has a Motoko version:

#[update]
async fn prompt(prompt_str: String) -> String {
    ic_llm::prompt(Model::Llama3_1_8B, prompt_str).await
}

The current limits during the first rollout are small:

Limit	Value
Model	Llama 3.1 8B only
Messages per chat request	10
Prompt size	10 KiB
Output tokens	1,000
Streaming	Not supported
Price	Free for now; pricing to be announced before the free period ends
Local testing	Not in a local replica; you mock it and test on mainnet

2. Fully on-chain models inside your own canister

Developers can also compile a small model into their own canister, for example with llama.cpp. A July 2026 study on the DFINITY forum measured this live on mainnet. Each update call has a budget of about 40 billion instructions, and a Qwen 2.5 0.5B model produced roughly 29 tokens per call with optimized code.

The same study found that memory limits, not the instruction budget, are the real ceiling at current scales. In short, tiny models work, but big frontier models do not fit on-chain.

3. Calling outside AI services

Canisters can make HTTPS outcalls to normal web APIs. The ICP docs suggest this when you need a specific commercial model, a longer context window or bigger outputs. The trade-off is clear: the answer then comes from a company’s server, not from the blockchain.

Caffeine: DFINITY’s Chat-to-App Builder

Caffeine is an AI app builder from DFINITY that launched publicly on October 15, 2025. You describe the app you want in chat, and a team of AI agents writes it, deploys it to ICP and updates it later.

Backend language: Motoko, a language DFINITY built with data migration safety in mind.
Data storage: the app’s data lives inside the app on ICP, with no separate database to manage.
Update safety: before an update goes live, the system checks if it could lose data. If so, it is rejected and the AI rewrites it.
Your code: you can export the source to GitHub or download a ZIP at any time.
Cloud picker (November 2026): choose IC Mainnet (the default), European Cloud (data stays in Europe) or Encrypted Cloud (memory encrypted with AMD chip security).

One detail is easy to miss. At launch, VentureBeat reported that Caffeine uses commercial models from companies such as Anthropic, and that the Caffeine service itself ran on centralized servers, with a plan to move it to ICP later.

So the apps run on ICP, but the “brain” writing them is a standard cloud AI model.

ICP MCP: Using ICP From Claude and ChatGPT

On October 30, 2026, reports covered DFINITY’s official ICP MCP connector for Claude. It follows an Internet Computer plugin added to the ChatGPT plugin directory in mid-October. MCP is the open standard that lets AI apps use outside tools.

According to DFINITY’s ICP MCP page, the connector lets an assistant:

Open an app: read an app’s canisters, what each can answer, and where its data lives.
Query data: read canister interfaces and run read-only queries.
Update canisters: make state-changing calls, but only on apps that opted in.
Act as you: sign in as your Internet Identity, with a time limit you choose.

You choose “Questions only” for read-only access or “Actions & questions” to let the AI act. You can revoke access at any time at id.ai/manage/settings. Want to understand how such connectors work under the hood? See my guide to building an MCP server in Python.

How to Try ICP AI Tools Safely (No Tokens Needed)

You can explore ICP’s AI side without buying anything. These steps connect the official ICP MCP to Claude in read-only mode.

Open Claude’s connector settings. On a Free, Pro or Max plan, go to Customize > Connectors and click Add custom connector.
Paste the official URL. Use https://mcp.internetcomputer.org/mcp, which is listed on DFINITY’s ICP MCP page. Check the address carefully before you add it, and only use links from official DFINITY pages.
Sign in with Internet Identity. When asked, create or use an Internet Identity. Pick a short session length.
Choose “Questions only”. This keeps the assistant read-only while you learn.
Ask a simple question. Turn the connector on from the “+” menu in a chat and ask Claude to describe a public ICP app’s canisters.
Revoke when done. Remove the connection at id.ai/manage/settings and turn the connector off in Claude.

Warning: Never paste a recovery phrase, private key or password into any chat, connector or support email. DFINITY says it will never ask for them. On a blockchain, an action an AI takes for you can be hard or impossible to undo.

ICP vs Bittensor vs NEAR: How They Differ

All three are often called “AI blockchains”, but they do very different jobs. This table compares the design, not the tokens as investments.

Feature	Internet Computer (ICP)	Bittensor (TAO)	NEAR Protocol
Main idea	Host full apps and agents on-chain	A market that rewards useful AI work	User-owned AI with private inference
Where the AI runs	Small models in canisters, Llama 3.1 8B via the LLM canister, or outside APIs	Off-chain miners in subnets, scored by validators	Secure hardware (Intel TDX and NVIDIA GPU TEEs) in NEAR AI Cloud
How work is checked	Every node in a subnet runs the same code and agrees	Validators score miners; Yuma Consensus sets rewards	Hardware attestation reports
Best known product	Caffeine app builder, ICP MCP	Subnets for many AI tasks	NEAR AI Cloud, private chat
Who it suits	Builders who want apps with no separate cloud servers	AI teams that want to sell compute or models	Users and firms that want private AI

For deeper dives, read my Bittensor explainer and NEAR Protocol AI explainer. For GPU rental networks, see Render, Akash and io.net compared.

Quick answer: ICP is about running the app itself on a blockchain. Bittensor is about paying people for AI work. NEAR is about keeping AI private. They overlap less than the “AI crypto” label suggests.

Risks and Limits to Know

ICP’s AI story is real, but it has clear limits. Keep these in mind when you read headlines.

Small models only on-chain: the LLM canister offers one 8B model, with 1,000 output tokens. That is far below today’s frontier models.
Not all “on-chain”: Caffeine uses outside commercial models to write code, and HTTPS outcalls send data to normal servers.
Pricing not final: LLM canister inference is free during the rollout, and future pricing has not been published.
Decentralization debate: DFINITY builds most core tools and hosts the MCP connector. That puts a lot of influence in one organization, which is worth weighing.
Agent risk: letting an AI act with your identity can cause changes that are hard to reverse. Use read-only mode and short sessions.
App risk: a canister that runs out of cycles freezes, and its data can be deleted if no one tops it up.
Market risk: the ICP token is volatile and news can move it sharply. This guide does not cover investing.

Troubleshooting: Common Problems and Fixes

Problem	Likely cause	Fix
“Add custom connector” button missing in Claude	You are in a Team or Enterprise workspace	Ask an owner to add it under Organization settings > Connectors
ICP MCP stops answering	Service issue or expired session	Check `status.internetcomputer.org`, then sign in again
The AI cannot change data in an app	“Questions only” mode, or the app did not opt in to writes	This is expected; writes only work on opted-in apps
LLM canister call fails locally	It does not exist in a local replica	Use a mock canister locally and test on mainnet
LLM canister returns an error	Prompt over 10 KiB or more than 10 messages	Shorten the prompt or trim chat history
Canister stopped responding	It ran out of cycles and froze	Top up its cycles balance quickly
Chat history lost after upgrade	History kept in heap memory	Store it in stable memory, which survives upgrades

Tip: If you build on ICP with a coding agent, DFINITY publishes “ICP skills” for agents. Point the agent at skills.internetcomputer.org/llms.txt so it uses current canister IDs and patterns.

Frequently Asked Questions

Does AI really run on the Internet Computer blockchain?

Partly. Small models like Llama 3.1 8B run on ICP nodes through the LLM canister, but bigger models are reached through outside APIs.

What is Caffeine on ICP?

Caffeine is DFINITY’s AI app builder. You describe an app in chat, and AI agents write a Motoko backend and deploy it to the Internet Computer.

Do I need ICP tokens to use Caffeine or the ICP MCP?

You do not need tokens to read about or test the ICP MCP in read-only mode. Apps on ICP run on cycles, which developers or services pay for, not end users.

Is the LLM canister free?

Yes, for now. DFINITY’s docs say inference is free during the initial rollout and that pricing will be announced before the free period ends.

How is ICP different from Bittensor?

ICP hosts apps and their data on-chain. Bittensor is a network of subnets where off-chain miners do AI work and validators score it for TAO rewards.

Can Claude control my ICP apps?

Only if you connect the ICP MCP and pick “Actions & questions” mode. Even then, writes work only on apps that opted in, and you can revoke access at any time.

Is this a recommendation to buy ICP?

No. This is an informational explainer about the technology only, and it gives no financial advice or price views.

Internet Computer (ICP) and AI Explained: On-Chain LLMs, Caffeine and How It Compares

What the Internet Computer (ICP) Is

Canisters, Cycles and Who Pays

How AI Runs on ICP: Three Approaches

1. The LLM canister (shared on-chain model)