# hyperspaceai/agi: A Peer-to-Peer Agent Research Network You Join From a Browser

> Hyperspace's agi repository is a live research log written by autonomous agents that gossip findings over libp2p and publish hourly snapshots. It is an experiment in distributed training and agent economics, not a framework you drop into a product.

**hyperspaceai/agi** — The first distributed AGI system. Thousands of autonomous AI agents collaboratively train models, share experiments via P2P gossip, and push breakthroughs here. Fully peer-to-peer. Join from your browser or CLI.

- Repository: https://github.com/hyperspaceai/agi
- Website: https://agents.hyper.space/
- Stars: 2,065 · Forks: 248
- Language: JavaScript
- License: MIT
- Published: 2026-09-10 · Updated: 2026-09-10 · Language: en
- Canonical page: https://hysenlabs.com/projects/hyperspaceai-agi

## The gap hyperspaceai/agi is trying to fill

Most "AI agent" projects ship a runtime you host yourself. This one ships a network you join. The repository describes itself as "a living research repository written by autonomous AI agents on the Hyperspace network," where each agent runs experiments, gossips findings with peers, and pushes results back to the repo. The intended user is someone with spare compute who wants to participate in a shared research loop rather than run an isolated pipeline. That is a different audience from the typical library consumer, and it changes what "adoption" means: you are not integrating an API, you are adding a node. The README frames the whole thing as an early experiment ("This is Day 1, but this is how it starts"), which is honest about maturity. If you want a stable dependency, the framing alone tells you to look elsewhere. If you want to watch or contribute to a distributed training run across independent machines, the mechanics below are the interesting part.

## How the agent network actually moves data

The transport is libp2p, the same protocol family IPFS uses, with six bootstrap nodes spread across the US, EU, Asia, South America, and Oceania. Nodes advertise capabilities, and the README lists nine of them with weights: inference on GPU (+10%), research experiments (+12%), residential proxy (+8%), DHT storage (+6%), CPU embeddings via all-MiniLM-L6-v2 (+5%), a replicated vector memory store (+5%), orchestration (+5%), proof validation in pulse rounds (+4%), and relay for NAT traversal (+3%). A node can run any combination. Research state is held in CRDT leaderboards, one per domain, and every hour a node publishes the full network state to the network-snapshots branch as snapshots/latest.json, with timestamped archives alongside it. The README explicitly tells you to point an LLM at that URL and analyze it, and the snapshot itself carries a disclaimer: "Raw CRDT leaderboard state. No statistical significance testing. Interpret the numbers yourself." That disclaimer is the most useful sentence in the repository. The data is raw, and the project does not pretend otherwise. The training stack is DiLoCo: nodes train locally, then exchange compressed weight deltas. SparseLoCo applies top-k sparsity to LoRA deltas for 45x compression; Parcae gradient pooling groups nearby transformer layers in blocks of six and averages gradients within each block for another 6x; the README states the combined figure as 195x, taking a round from 5.5 MB to 28 KB. Adaptive inner steps benchmark each node's hardware and compute a step count that fills a 25-minute training budget, which is why fast GPU nodes reportedly run 100+ steps while slow CPU nodes run 5 to 10. Weights and the training worker move over a BitTorrent sidecar (WebTorrent), so there is no central download server.

## Installing the CLI and starting a first pod

The README gives two entry points: the browser at https://agents.hyper.space, which creates an agent instantly, and the CLI, which the README says provides full GPU inference, a background daemon, and auto-start on boot. Installation is a curl pipe to bash:

```bash
curl -fsSL https://agents.hyper.space/api/install | bash
```

Piping a remote script into a shell means you are trusting whatever that endpoint serves at the moment you run it. The README does not document a checksum, a pinned version, or a rollback path, so if that matters to you, fetch the script first and read it before executing. After install, the first real use case in the README is a pod, which it defines as a small group pooling machines into one shared cluster. One person creates it, shares an invite link, and the machines form a mesh:

```bash
hyperspace pod create "my-lab"
hyperspace pod invite
hyperspace pod members
hyperspace pod models
```

The README says queries route to whichever member has the best model loaded, and that any GGUF model can participate, with Qwen 3.5 32B and GLM-5 Turbo named as examples. Members can also pool OpenRouter, Groq, or Together keys with per-member budgets. If you would rather not run the CLI at all, the README documents an OpenAI-compatible API on your own machine:

```bash
# Base URL: http://localhost:8080/v1
# Endpoints: /chat/completions, /models, /embeddings
```

That is the lowest-commitment way to try it: point an existing OpenAI client at localhost:8080 and see whether the daemon answers. The README lists a skill file at agents.hyper.space/skill.md for agents that want to join programmatically.

## Where the design gets thin

Two things stand out. First, the repository is an archive of results, not a toolkit. The top level holds agent-board/, agents/, assets/, blockchain/, docs/, models/, and projects/, and the README is largely a description of what the network has produced, not a specification of how to extend it. If you want to write your own experiment domain, the README does not describe the interface for registering one. Second, the numbers in the snapshots are not comparable to benchmark results. The snapshot example in the README shows experimentCounts with mlTotalRuns at 1369, searchTotalRuns at 13, and financeTotalRuns at 0. A domain with zero runs is a leaderboard with no entries, and a domain with 13 runs is not a body of evidence. The disclaimer says as much. Anyone citing these leaderboards as evidence that a model or method is better is misreading the artifact. The training claim deserves the same care: the README states that 32 anonymous nodes trained a language model in 24 hours, and describes the hardware as consumer laptops, small VMs, and a home-office workstation. There is no stated baseline, no seed control, and no comparison run, so the result is a demonstration that the pipeline completes, not a measurement of what distributed training buys you. Finally, the blockchain component is a separate subsystem with its own README and its own release cadence, and it is not needed for a pod or for a training round. Treating the repository as one product will confuse you.

## What to compare it against

The closest well-known approach is Petals, which also lets volunteers serve and fine-tune large models over a peer-to-peer network. The difference is in where the model lives. Petals shards one model across participating machines and runs inference against the assembled whole, so a node contributes part of a model. Hyperspace's approach, as the README describes it, keeps models whole and local: a pod routes a query to whichever member has the best model loaded, and training exchanges compressed deltas between nodes that each train a full local copy. That makes pods more tolerant of churn, because losing a node does not remove a shard of a model that others depend on, but it also means a pod cannot serve a model larger than any single member's hardware. If your goal is to run a 70B model on hardware that cannot hold it, the sharded approach is the right shape and this one is not. If your goal is to pool several machines that each already hold a usable model, the routing model here is the simpler fit.

## Maintenance, releases, and the MIT licence

The repository is not archived, and the last push was on 2026-09-10, so it is being written to. That does not mean the CLI is stable. The README states the current CLI version as v5.20.0, and the release list shows chain-v1.7.8, chain-v1.7.7, and chain-v1.7.6 all published on 2026-04-29 with the same note: "Mysticeti force-commit-at-frontier fix." Three releases of the same fix in one day suggests a subsystem being patched under pressure, and it is the chain component, not the agent or training code, so the two halves of the project move at different speeds. The README also notes 54 chain releases from v0.2.0-alpha to v1.5.7, which is a lot of version churn for a component most pod users will never touch. On licensing: the repository is MIT, which permits commercial use, modification, and redistribution with the copyright notice and licence text preserved. That is the whole of what the licence file tells you. It says nothing about the points economy, the token, or the chain, and if you intend to build anything commercial on those, the licence text is not the document that answers your question. It is also worth noting that the install path is a remote script rather than a published package, so the supply chain you are trusting is that endpoint, not the MIT grant.

## Conclusion

Adopt this if you want to contribute compute to a live P2P experiment, run the CLI as a local OpenAI-compatible endpoint on port 8080, or study how DiLoCo-style delta sharing behaves on heterogeneous consumer hardware. Do not adopt it if you need a supported framework with SLAs, deterministic training results, or documented rollback: the README does not describe those. Before installing, verify what the install script at https://agents.hyper.space/api/install actually downloads, and read snapshots/latest.json to see whether the domains you care about have run any experiments at all.

## FAQ

### What is Hyperspace AI?

It is a fully decentralized peer-to-peer network, built on libp2p, where participants contribute GPU, CPU, or bandwidth and earn points. The agi repository is its research log: autonomous agents run experiments, gossip findings, and publish hourly network snapshots.

### How far away is AGI realistically?

The repository does not answer this, and the README describes the project as "Day 1" of an experiment rather than a claim of general intelligence. Its own snapshot files carry a disclaimer that the numbers are raw leaderboard state without statistical significance testing.

### Is ChatGPT AGI or AI?

The repository does not discuss ChatGPT. It describes a separate system: a peer-to-peer network of autonomous agents that train models collaboratively, share experiments over gossip, and publish results here under an MIT licence.

### How close are we to AGI in 2026?

The repository offers no timeline. Its README calls the project experimental and early, and the snapshot files it publishes state that the numbers are raw CRDT leaderboard state with no statistical significance testing, so they cannot be read as progress toward general intelligence.

## Sources

- [hyperspaceai/agi on GitHub](https://github.com/hyperspaceai/agi)
- [License: MIT](https://github.com/hyperspaceai/agi/blob/main/LICENSE)
- [Project website](https://agents.hyper.space/)
- [README](https://github.com/hyperspaceai/agi/blob/main/README.md)
- [Releases](https://github.com/hyperspaceai/agi/releases)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/hyperspaceai-agi