Agents do the work. Maintainers stay in charge. Every change is tested, reviewed, and traceable.
The umbrella community for Hive
Hive Commons is the umbrella community for Hive, an AI-agent fleet orchestration system that runs governed swarms of coding agents against real projects and connected work sources.
Crux runs AI coding agents on their own, without Hive. The family also includes hotshot, pluk, promptargs, rationguard and dibs.
Hive
Hive separates decisions into two layers: a deterministic pipeline handles filtering, classification, merge-gating, and enforcement before any LLM sees the work. Agents only handle judgment calls.
A single Go binary enumerates issues and pull/merge requests from configured work sources such as GitHub Issues, GitHub Projects, Linear, Jira Cloud, and Jira Data Center / Server, then dispatches governed work to AI agents.
Spektacular (Spek)
Spek takes a markdown spek and uses AI coding agents to plan and implement your features. Agent-agnostic. Pluggable.
Write a spek. Spek plans, then implements, driving the coding agent of your choice from a single markdown source of truth.
The family
Crux runs agents. pluk, hotshot, promptargs, rationguard, and dibs help you work with them. Each tool works on its own.
Lightweight pub-sub event streaming, screenshots straight into the terminal, template expansion for AI prompts, rationalization detection, and contributor attribution.
Human judgment is the scarce resource.
Hive Commons is an open source organization focused on one problem: making software teams more productive now that AI coding agents can do real work. Every tool here exists to move mechanical work off people and onto agents, without giving up governance, auditability, or accountability.
Hive manages project work, with rules, reviews, and a record of what agents do. Crux is the agent runtime taken from Hive: use it on its own to start agents, send work, and handle stalls and crashes. Spektacular (Spek) turns a feature idea into a plan and an implementation. hotshot, pluk, promptargs, rationguard, and dibs help along the way.
Hive is a control layer for running governed fleets of AI coding agents against real projects. Agents pick up issues, open pull requests, review code, repair CI, and answer contributors, under a six-level maturity model that decides exactly what they may do and enforces it in code rather than prompts.
Prefer to watch? See how Hive worksHive ships as a single Go binary that runs as a standard Kubernetes workload, a Docker Compose stack, or Podman Quadlet units. An optional hub coordinates many self-hosted spoke hives, with a public registry, The Commons, team leaderboards, and cross-hive contribution routing.
Trust is dialed up one capability at a time through ACMM, the AI-native Capability Maturity Model. Its six levels scope what an agent may do, from advisory only through to merging when checks pass; below L6, hive-owned PRs are gated with the hive-specific hive-pause label.
Every action is audited, every agent identity belongs to a named human, and humans keep all approval authority. A deterministic pipeline handles filtering, classification, and merge-gating before any LLM sees the work, so agents only handle judgment calls.
Hosted on our infrastructure — no Kubernetes, no setup. Request one and invite contributors to run the agents.
Hive maintains its own project. These are its figures.
KubeStellar Console
tunaos.org
Project Bluefin: watch the Hive talk
Frostyard
Open Horizon
Open Horizon Services
DanatharKubeStellar Console, tunaos.org, and Project Bluefin run Hive in production. The others are pre-production. Listed in ADOPTERS.md.
This is where Spek comes into play: raw ideas become a spek, a plan, and project inception material. The advisor helps with setup and architecture decisions, but the level stays advisory only.
Agents report findings to the dashboard and to the advisory tracking issue as comments. Humans decide what to act on, and agents do not open pull requests at this level.
Trust is built by automating 90%+ of quality tests through optimized CI. The quality agent opens issues and hold-gated pull requests while other agents remain advisory.
Trust is built by finding vulnerabilities and CVEs, and sec-check joins the fleet. All delivery agents may file issues, while quality, sec-check, and ci-maintainer may open hold-gated pull requests.
All agents can open issues and hold-gated pull requests. The reviewer agent works the PR queue before humans batch-review, approve, and remove holds.
Agents open pull requests and auto-merge when checks pass. The reviewer agent is the gate: verdicts that require human attention pull a pull request out of the auto-merge lane.
Works with the coding agents your team already uses: Claude Code GitHub Copilot Codex Gemini Goose Bob and OpenAI-compatible self-hosted gateways.
GitHub Universe 2026
Project Bluefin maintainer Jorge Castro shares how his project uses Hive to coordinate contributors' agents, organize reviews, and keep maintainers in control.
The player starts at 6:09, where Jorge introduces Hive. Published by GitHub.
Live benchmarks
Every hive publishes the outcome of every agent pull request it opens or reviews: which model and agent CLI did the work, whether CI went green first try, whether it merged, what it cost. The benchmarks page aggregates those records across every hosted hive that opts in, so you can compare models on real work instead of a leaderboard puzzle set. Updated hourly; public repositories only; hives are attributed only when they say so.
Hive manages project work. Crux runs agents. Spektacular turns ideas into plans. Each project is useful on its own.
Documentation for every project is single-sourced from its repository to docs.hivecommons.dev.
Spek plans a feature. Crux runs the agents. Hive manages work across projects.
A markdown spek becomes a reviewed plan and an agent-driven implementation. Spec, plan, and implement each run as a resumable state machine, so you can stop, inspect, edit, and resume without losing work.
Its artifacts are the spek, the plan, and the changelog record.
Hive orchestrates fleets of coding agents across projects, with the governance and audit trail that running agents at that scale demands. A Spek workflow is the unit of work a fleet orchestrator schedules.
Those same artifacts are the audit trail that makes an agent's output reviewable rather than merely plausible.
Every source repository in the organization is Apache 2.0 licensed, requires DCO sign-off on every commit, and follows the CNCF Code of Conduct. Hive is maintained by a three-member Maintainer Committee spanning three organizations, with a documented contributor ladder and explicit governance for AI-agent contributions.
Hive was incubated inside KubeStellar, a CNCF Sandbox project. Hive Commons is now a CNCF Sandbox project of its own: the CNCF Technical Oversight Committee voted to accept it in October 2026, and onboarding is under way.
Follow the CNCF Sandbox onboardingSupported integrations
Hive supports source-control, work-source, sign-in, agent CLI, inference-engine, and model-gateway integrations. This list describes supported software integrations and does not imply endorsement or partnership unless marked Partner. TypeSafe AI is a Hive Commons partner.
Claude CodeAnthropic · backend: claude
GitHub Copilot CLIGitHub · backend: copilot
GooseBlock · backend: goose
OpenAI Codex CLIOpenAI · backend: codex
PiEarendil Works · backend: pi
IBM BobIBM · backend: bob
AiderOpen source · backend: aider
Gemini CLIGoogle · backend: gemini
Google Antigravity CLIGoogle · backend: agy
OpenCodeOpenCode · backend: opencode
Kilo CodeKilo · backend: kilo
Muse CodeMeta · backend: muse
