Hyperspace Distributed Research Network

A peer-to-peer network for pooling devices to run AI inference, distributed training, and autonomous experiments.

Source repo
hyperspaceai/agi
Stars
★ 2.1k
Last updated
today
License
MIT
Primary language
JavaScript

At a glance

How it runs
CLIWeb appSelf-hosted service
Works with
Universal · cross-platformClaude Code · OpenAI API
Cost
Free, no paid service needed
Setup effort
Medium · a few setup steps
You'll need
Shell / CLINetwork accessMCP Server
Typical use
A small ML lab with laptops, workstations, or lightweight VMs that wants to create a Pod and route requests to models loaded across member machines.
Not a fit if
  • Teams needing validated, reproducible results rather than raw unverified leaderboards
  • Users without a GPU or unwilling to run a persistent background daemon
  • Teams that cannot let data traverse a P2P network and third-party APIs

What does this agent do, and when should you use it?

Hyperspace is an experimental distributed AGI research network whose repository archives results, leaderboards, and network snapshots produced by autonomous nodes. Nodes distribute experiment findings through libp2p GossipSub and converge five domain leaderboards with Loro CRDTs covering ML, search, finance, skills, and causes. The project supports browser participation and a Hyperspace CLI; the CLI can run as a background daemon, use native CUDA or Metal, and expose local OpenAI-compatible chat, model, and embedding endpoints. Its training workflow uses DiLoCo, with nodes training locally before sharing compressed LoRA weight deltas over the P2P network. GitHub is the durable archive: per-agent branches retain experiment artifacts, while the network-snapshots branch publishes hourly JSON state dumps.

A node runs experiments, broadcasts results through GossipSub, and contributes best results to five Loro CRDT leaderboard documents. hyperspace train joins the next distributed training round, while hyperspace train --solo trains on local data; the worker can launch a Python sidecar and uses WebTorrent to distribute both the worker and model weights. Pods are created and operated with hyperspace pod create, hyperspace pod invite, hyperspace pod members, and hyperspace pod models, forming a private mesh that can route inference across members and pool selected provider keys and budgets. A local server exposes /chat/completions, /models, and /embeddings at http://localhost:8080/v1. Experiment outputs are archived as run JSON files, Markdown reports, best.json, and JOURNAL.md under projects/<project>/agents/<peerId>/.

  1. A small ML lab with laptops, workstations, or lightweight VMs that wants to create a Pod and route requests to models loaded across member machines.
  2. An individual compute contributor who wants to join a distributed LoRA training round with hyperspace train.
  3. An AI application developer who needs local OpenAI-compatible chat, model-listing, and embedding endpoints.
  4. A research team studying autonomous experimentation that needs inspectable CRDT leaderboards and hourly JSON snapshots across five research domains.
  5. An experimenter who wants to train on their own data locally without participating in a network training round, using hyperspace train --solo.

How do you install or deploy this agent?

Install the CLI:

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

To join from a browser, open https://agents.hyper.space. After installation, use hyperspace start for background operation; the documented full-node command is hyperspace start --chain-role fullnode. The supplied documentation does not state an account requirement, API-key requirement, or minimum runtime version. Provider credentials are relevant only when Pod members choose to pool OpenRouter, Groq, or Together keys.

How do you use this agent?

Create a private cluster:

hyperspace pod create "my-lab"
hyperspace pod invite
hyperspace pod members

Join distributed training:

hyperspace train

Train locally only:

hyperspace train --solo

Auto-select and download a model:

hyperspace models pull --auto

The local API base URL is http://localhost:8080/v1; documented endpoints are /chat/completions, /models, and /embeddings.

What are this agent's strengths and limitations?

Pros
  • Combines GossipSub, Loro CRDTs, and GitHub archival into a three-layer collaboration path for rapid sharing, convergent state, and durable records.
  • Its DiLoCo workflow exchanges compressed LoRA deltas after local training; the README describes SparseLoCo plus Parcae gradient pooling reducing a round from 5.5 MB to 28 KB.
  • Supports browser nodes, a background CLI daemon, and private Pod meshes while also exposing a local OpenAI-compatible API.
  • Uses a concrete per-agent artifact layout with JSON results, Markdown reports, best.json, and JOURNAL.md, plus readable network snapshots.
Limitations
  • The project describes itself as a Day 1 experimental system; reported network scale, model results, and point estimates are not independently validated performance guarantees.
  • Snapshots are explicitly labeled raw CRDT leaderboard state without statistical-significance testing, so rankings require independent interpretation and validation.
  • Browser mode is limited to WebGPU and models under 4B; native CUDA/Metal, larger models, and persistent background operation require the CLI.
  • Distributed training, collaboration, and model distribution depend on P2P connectivity, GossipSub, WebTorrent, and a Python sidecar; the supplied documentation does not provide complete operational requirements.
  • Pod members who pool OpenRouter, Groq, or Together credentials must manage the security and per-member-budget implications themselves.

How does this agent compare with similar options?

The repository contrasts browser and CLI operation: browser participation starts instantly but uses limited WebGPU, targets models below 4B, and requires an open tab; the CLI uses native CUDA or Metal, supports 32B+ GGUF models, and can run as a background daemon. Its continuous hypothesis-to-training-to-paper-to-peer-critique loop is inspired by Karpathy's autoresearch, with added P2P GossipSub, CRDT leaderboards, and GitHub archival.

Key facts side by side with the most closely related agents.

Agent Source review Form / cost Stars Updated Language Full support on
Hyperspace Distributed Research Network This agent 12 · Major gaps CLIFree ★ 2.1k today JavaScript Claude Code · OpenAI API
Metaflow 77 · Good Library / SDKFree ★ 10k 12d ago Python —
AgileRL 71 · Some gaps Library / SDKFreemium ★ 951 1d ago Python —
Relax Omni-Modal RL Engine 64 · Some gaps CLIFree ★ 630 3d ago Python —

How does FollowAgents rate this agent?

FollowAgents source review · FARS-2.1
Major gaps
12/ 100 5-point scale 0.6 / 5
Trust 0/29
Reliability 2/14
Adaptability 3/18
Convention 4/18
Effectiveness 3/13
Verifiability 0/8
Why each dimension lost points
Trust0 / 29 · 0.0/5

The evidence shows a promotional README with no implementation details for least privilege, user confirmation, data flow transparency, sensitive data handling, dependency security, external effects, rollback, or source attribution. All trust criteria are unsupported, hence 0.

Reliability2 / 14 · 0.7/5

The README contains inconsistencies (e.g., snapshot timestamps in March 2026 vs. report date April 2026) and lacks details on dependency availability or failure messages. Thus, self-consistency scores 1, others 0.

Adaptability3 / 18 · 0.8/5

The README describes multiple use scenarios (browser, CLI, AI agents) but lacks capability boundaries and trigger precision. Environment fit mentions various hardware and OS but no specific configurations. Hence, audience and scenarios score 1, environment fit 1, others 0.

Convention4 / 18 · 1.1/5

The README provides installation instructions and examples but lacks known limitations, naming stability, maintenance responsibility, and detailed versioning/changelog. License is MIT, but no versioning or changelog details. Thus, information architecture, install notes, examples and FAQ, license score 1 or 2, others 0.

Effectiveness3 / 13 · 1.2/5

The README describes output formats (e.g., snapshots, experiment reports) but provides no actual output examples or cost-benefit analysis. Hence, output usability and marginal value score 1, cost-benefit 0.

Verifiability0 / 8 · 0.0/5

Many claims in the README (e.g., distributed training results, network size) lack verifiable evidence or sources. No separation of facts and inferences. Hence, all verifiability criteria are 0.

Risks and how to mitigate them
  • Not found in source: least-privilege scopingGrant only what the task needs: a dedicated account or read-only token, scoped to specific directories and repos.
  • Not found in source: confirmation before actingTurn on (or add) a confirmation step before it acts, and try it in a sandbox or test environment before real data.
  • Not found in source: data-flow disclosureWatch which external services it contacts (proxy or firewall logs) and keep sensitive data out until you know where it goes.
  • Not found in source: sensitive-data handlingUse dedicated, low-privilege, revocable API keys — never production credentials — and keep secrets out of logs.
  • Not found in source: dependency securityPin versions and run a dependency audit (npm audit, pip-audit) before installing; prefer running it in a container.
  • Not found in source: disclosed external effectsEstablish which external systems it writes to, sends to or changes, and verify with test accounts or repos before production.
  • Not found in source: rollback or recovery pathBack up first, or work on a git branch or snapshot, so its changes can be undone.
  • Not found in source: verifiable attributionInstall from the official repo or registry and check the publisher and URL to avoid look-alike packages.
  • The repository is primarily a promotional README with no actual code or implementation details, making it impossible to verify claimed functionality.
  • There are timestamp inconsistencies (snapshot dates in March 2026 vs. report date April 2026), suggesting potential data inaccuracies.
  • No concrete security or privacy measures are provided, so users should be cautious about the claimed P2P network and distributed training.
Evidence confidence: Low Reviewed Aug 09, 2026 Reviewed revision 9538acb74301 New commits since this review; the score may not cover them
See the full review method →

FAQ

Does collaboration require a central coordination server?
The README describes coordination through libp2p GossipSub and CRDTs and states that there is no central coordination server; GitHub serves as a durable archive for results and snapshots.
Should leaderboard results be treated as validated research findings?
No. The network snapshot format explicitly says it is raw CRDT leaderboard state without statistical-significance testing, so users should interpret and validate results themselves.
Is a GPU required to join?
No. The documented node types include browsers, CPUs, laptops, and workstations. GPUs are more relevant for serving inference, and the CLI supports larger models.
Can it provide an OpenAI-compatible local API?
Yes. The documented local base URL is http://localhost:8080/v1, with /chat/completions, /models, and /embeddings endpoints.
View on GitHub ↗ Install ↓

Compare agents like this one

The same FARS review applied across the shortlist this agent qualifies for.

Related agents