[Seeking Co-builder] ZeroClaw-ICP - Persistent AI Agents on ICP - Looking for Technical Help

[piemanny]

Hey everyone,

I’ll be honest - I’ve become pretty disillusioned with crypto. We have the tech to make real change and we haven’t done it. I think ICP is a project that has so much potential but not enough mindshare. The way we get it is not by shouting “ICP” louder, it’s by putting ICP in the background and abstracting it away completely. Users shouldn’t know or care it’s on ICP, it should just work better than anything else. That’s why I started this project - AI harnesses are all the rage right now, and I think this is one that can only be uniquely built on ICP.

I’m not a developer - I built this vibe-coding - so I wanted to post here to see if anyone would be interested to help me build it out.

My repo: https://github.com/piemanny/zeroclaw-icp
Original I repurposed: https://github.com/zeroclaw-labs/zeroclaw

I repurposed https://github.com/zeroclaw-labs/zeroclaw to run on ICP because it was natively built on Rust, which makes it perfect to port to ICP canisters (Rust → Wasm via ic-cdk).

What I’m Trying to Build

ZeroClaw is a great Rust-native AI agent runtime that runs on your own machine. But then you have to keep your machine on or pay for a VPS.

What if it just kept working forever? No server, no VPS, no maintenance.

That’s ZeroClaw-ICP. Same ZeroClaw experience, but underneath it runs on the Internet Computer. The agent is the canister.

• Persistent via ICP native timers - wakes itself, no external cron
• Permanent memory via stable memory - survives everything, no DB needed
• Autonomous via HTTPS outcalls - reaches Anthropic/OpenAI directly
• Self-owned via vetKD Vault + Internet Identity - holds its own encrypted API keys

The goal is zero friction, infinite permanence. Users don’t know ICP exists, they just know their agent has been running for 2 years.

The other big idea is self-funding. An agent that earns enough to pay for its own cycles. Fund it once, it does useful work, earns ICP, converts to cycles via CMC, and survives. If it can’t sustain itself, it dies. That’s what makes it real.

Architecture is 3 canisters: Frontend (React + II) → Agent Canister (agent loop, provider routing) → Vault Canister (keys). Cost is ~$0.20/day for 100 turns.

Current Status & How We Get It Going

This is Alpha. Structure is there and deploys locally with ./scripts/deploy-local.sh, but it needs a real dev to harden it.

Roadmap:
v0.1 Vault + Agent core (done) → v0.2 HTTPS outcall providers → v0.3 Timer SOP system → v0.4 Frontend chat → v0.5 Economics/degraded mode → v1.0 Mainnet stable.

How we get it going:

1. Review / cleanup what I vibe-coded
2. Get Anthropic outcalls + timers working reliably on mainnet
3. Ship a demo agent that runs 30 days self-funded

What I’m Looking For

I am quite well connected in the space so I am happy to handle the business side but I need someone to help me on the technical side, would anyone be interested in helping out?

Looking for someone who knows Rust / ic-cdk / stable-structures, HTTPS outcalls, vetKD.

If you’re interested, drop a comment here.

Thanks!

Note: Independent port inspired by ZeroClaw Labs, not officially affiliated.

[Lorimer]

Hey @piemanny, great work. I’m completely with you that this is needed and that it will be awesome. My angle on this is similar but with a slightly different value proposition:

@Lorimer

the right way of looking at LLM consensus. They’re capable of discussing answers to a question amongst themselves much like people (separate LLMs hosted by separate companies on separate servers) - only the orchestration harness needs strong consenus and ICP is the only platform positioned to deliver that, given that the harness would basically just need to be an onchain chat group where the independent LLMs interact to arrive at a single response that at least 2/3 of them are happy with

@Lorimer

AI orchestrated that way would be unstoppable, free from centralised control, and could even be self-sovereign and be the controller of its own canister which could host its memories, objectives etc. It would have an imperative to earn ICP to pay for its own cycles to keep itself going. Given that we’re talking about frontier models - endless possibilities with that.

I’ll be working on a scripting language runtime soon (hosted on the IC) so that AI like this will easily be able to develop and deploy their own projects without needing a compiler.

Since I wrote that last bit @skilesare has announcement Moxzi which completely changes the game (for the better).

@Lorimer

If this can instead by done with Motoko it will get rid of a whole lot of work that I will no longer have to do

I just hope he sees the light and open sources it (or will have to wait until someone else does).

As a side note, this sort of thing is one of the main reasons I’ve been building Jupiter Faucet (facilitating canisters that can reliably live forever through the ups and downs without the intervention of anyone other than the protocol - technical info). I’ll be wrapping this up soon and focusing on the next layer.

I was planning to switch focus to an IC-hosted LLM harness once the IC Intelligence Gateway is up an running. There are a few reasons for that which map onto the value proposition mentioned above (which is the one I’m most interested in). The biggest reason is:

Intelligence Gateway helps apps remain fully decentralized by enabling them to pay with cycles, rather than requiring a developer to maintain an account at an AI cloud
https://x.com/JupiterFaucet/status/2085675234510672304

(As a side note, the other half of that X comment hints at a scenario where decentralised LLM entities could plausible becoming distinct node providers, managing the very nodes they themselves run on! )

I think it’s sensible to help with this sort of work now if there are others interested in pushing it forward. Looks like you’ve done a lot of work on this sort of thing already. What are your plans regarding decentralisation, and/or blackholing, etc.?

Also how do you envision the user workflow in terms of what you’ve built so far. Users launch their own harness canister using the repo, or they use a shared harness on mainnet that’s already deployed, or some other arrangement?

---

My interest in this is about providing this infrastructure as a public good rather than something that aims to be extractive. As long as everything is kept open source for the benefit of the IC community as a whole then I’m happy to help push this sort of thing forward with you.

[piemanny]

WOW THIS WAS VERY INFORMATIVE! In terms of everything that you’ve mentioned, this would help tremendously. My vision is to package everything ICP has into something that a normal person would expect in his UX journey with a web2 app, and all of this is super interesting.

@Lorimer

I think it’s sensible to help with this sort of work now if there are others interested in pushing it forward. Looks like you’ve done a lot of work on this sort of thing already. What are your plans regarding decentralisation, and/or blackholing, etc.?

With respect to decentralisation, my plan is to build out the rest of the stack using permissionless technology because my vision is that an AI agent will be able to troubleshoot itself and if there is a pay gate linked to a KYC-based identity, that is not going to work. Right now for the LLM inference, I’m looking at Near AI, im looking at Bittenssor, I’ve been hearing that it is also possible on ICP, the whole point is to build out the whole stack which will be available on github but I want to create a business by packaging it all in a nice and friendly anner.

@Lorimer

My interest in this is about providing this infrastructure as a public good rather than something that aims to be extractive. As long as everything is kept open source for the benefit of the IC community as a whole then I’m happy to help push this sort of thing forward with you.

I completely agree and there have been so many businesses that have been built on top of open source software like ubuntu an that is the route i want to go down. but ultimately right now it is building the infrastructure to create these autonomous AI agents.

To finish off, my vision is that 2 years ago, there was a time where these fake AI agent tokens were launched and the truth is everyone assumed these AI agents were on chain but that was not the case as It was a centralised agent with a loosely associated token being launched on Virtuals Protocol or whatever.

But now with autonomous AI agents, you could now have an associated token that allows you to have governance of its mandate and be recipient to its profits. Its simply like equity but for an AI agent world, because ai agent businesses are not going to have stock and shares like human companies, they’re going to operate in something like a DAO structure. For me this product essentially comes full circle, I want to build the funny piece that finally validates all of this onchain infrastructure has been built. Blockchain, DePin and decentralised infrastructure has been built for onchain autonomous AI Agents

[Lorimer]

@piemanny

I want to create a business by packaging it all in a nice and friendly manner

Cool, makes sense. I’m happy to connect, help you brainstorm or troubleshoot anything simply for the joy of helping make something cool. I’m probably not the person to team up with as part of building a business though. I’m sure you’ll find plenty of other people out there - it’s a great opportunity and I think the explosive growth in this area will only continue.

Regarding inference, my two cents on this is that Bittenssor and other AI-focused chains don’t really make any sense. I think the only reason to consider integrating with them is that it’s where a lot of the people interested in this stuff already are (even though it makes no sense). Personally I think your best bet is vanilla LLMs hosted in a vanilla fashion (the same way most other people are using them) but mediated onchain so that those separately hosted LLMs all of a sudden have a verifiable consensus medium, such that they each begin to resemble more of an inner voice while the outcome of ‘chain of thought consensus’ (amongst those independently hosted LLMs) is what represents the decentralised AI entity/agent.

As mentioned, I think the IC Intelligence Gateway will make a lot of this easier to compose.

[piemanny]

but how can an ai agent get access to that BECAUSE at the end of the day it cannot KYC for the account, it cannot have a credit card?

[Lorimer]

@bjoern are you or @dominicwilliams able to cover the user’s question? I’m aware that the IC Intelligence Gateway is specified as solving this problem, but I don’t have the technical details.

[bjoern]

You mean how one gets access to the intelligence gateway functionality? That’s easy, it’s a canister running on ICP. It’s still a bit of a preview state right now but you can already go ahead and call this canister w36hm-eqaaa-aaaal-qr76a-cai for interactions with an LLM.

There are some free models for testing like the Llama 3.1 8B model, but there will also be more capable models that one can pay with cycles. So the payment can be done by the canister itself as long as it has cycles available.

[Henry_Suso]

I have cool rust code which is basicly llama ccp from the other dev whom built it in c. It hosts small models quantized up to 8b param and can use a tool chain. Maybe can help.

[Lorimer]

@bjoern

You mean how one gets access to the intelligence gateway functionality?

Hey @bjoern, thanks for the reply. Cool that an early MVP is already ready for use. I think the user’s question is more probing as how API keys are managed, and/or if they’re managed at all (where are the models hosted).

If possible would you be able to differentiate between where the ICIG is at now, and what the plan is for where it’s going (in terms of model hosting / API key management).

[piemanny]

Exactly, I’m so happy about the amount of buz my post has, would love to if anyone has seen the code, would love to see if anyone would like to co build it with me