A Claude Code plugin — Open source

Your AI team.
Actually a team.

Hive turns a Claude Code session into a coordinated swarm of 25 AI specialists — with the structure, gates, and memory of a real software organization.

/hive:kickoff
north star set · roster online
/hive:plan "add checkout flow"
epic · 7 stories · dependency-tracked
/hive:execute — swarm at station…
The roster

A swarm with roles — not one model improvising.

Twenty-five specialized personas, each with a defined job, coordinating through structured workflows.

01

A swarm with roles

Analyst, architect, developer, tester, reviewer, and more — each a defined station, coordinating through structured workflows instead of one model doing everything.

02

Discipline built in

Every project moves through kickoff, plan, execute, review, test, and ship. Human-in-the-loop gates keep you in command at every phase boundary.

03

Cross-model execution

Route implementation to OpenAI Codex while Claude orchestrates and reviews. Two models, adversarially paired — lower cost, and no model grading its own homework.

04

The test swarm

A five-agent pipeline authors tests, runs them, files bugs, and routes fixes automatically. Failures become tickets, not surprises.

05

Memory that compounds

Layered L0–L3 memory persists decisions across sessions and projects: session insights, a compiled wiki, a knowledge graph, and an optional semantic index.

06

Extensible to the core

Add agents, skills, workflows, and whole teams without touching Hive's core. It's a substrate, not a walled garden.

A hive of specialist agents — a honeycomb lattice with cells lit amber as work flows between them
The pipeline

Six phases. Human gates between them.

01

Kickoff

Discover the codebase or shape the greenfield vision; set the north star.

/hive:kickoff
02

Plan

Decompose requirements into an epic of dependency-tracked stories.

/hive:plan
03

Execute

The swarm implements stories in parallel, each agent at its station.

/hive:execute
04

Review

Structured, adversarial code review — findings become revision loops.

/hive:review
05

Test

The five-agent test swarm exercises the build and files what it finds.

/hive:test
06

Ship

Reconcile, version, release — with a human holding the final gate.

/hive:ship
// Interop

Native to Claude Code.
Fluent in everything else.

Hive installs as a Claude Code plugin and runs with zero extra infrastructure — but it was never a walled garden. It's a coordination layer that treats agents and harnesses as pluggable parts.

  • Claude Code native. One plugin install. The lead agent describes a team; the runtime spawns it.
  • OpenAI Codex, out of the box. Cross-model by design — Claude orchestrates and reviews, Codex implements. One key in agent_backends gives any agent a different brain.
  • Extensible by contract. New harnesses, new agents, new skills — added at the edges, without touching core.
Two luminous hex nodes exchanging a braided stream — Claude orchestrating, Codex implementing, each fanning out to agents
// Execution substrates

You choose the engine. Hive is the transmission.

The same workflows, plans, and memory run on any substrate. Swap engines in hive.config.yaml or with HIVE_EXECUTION_MODE — your process doesn't change, your throughput does.

claude-directthe default

Zero infrastructure

The lead describes a team in natural language and the Claude Code runtime auto-spawns the teammates, right in your session.

USE WHENStarting out, solo, or shipping a normal-sized feature. Nothing to install.
claude-workflowsstructured dag

Workflow-driven on Claude

Run the classic workflow DAG — dependency-tracked stories driven through defined phases by Claude agents, no external daemon.

USE WHENYou want structured, phase-gated, repeatable runs without standing up a meta-harness.
multicathe meta-harness

DAG execution substrate

Issue- and task-driven stories, parallel git worktrees, a managed daemon keeping the swarm honest.

USE WHENBig epic, parallel stories, managed & resumable runs. Opt in with /hive:multica-init.

Three engines today. The substrate contract is open — bring your own tomorrow.

// Layered memory

Your agents stop forgetting.

Most agent teams wake up with amnesia. Hive's memory is layered — each level optional, each compounding on the last. Session insights become team knowledge; team knowledge becomes queryable provenance. Ask /hive:why and the knowledge graph answers with the decision, the context, and the chain of reasoning behind it — no git archaeology. And memory travels across projects, so lessons from one product make the next one faster.

Four translucent hexagonal memory layers stacked in depth — L0 session to L3 semantic index — joined by a rising signal thread
L0SESSION

Per-session insights, harvested as agents work. Free, automatic.

L1TEAM WIKI

Compiled team memory — conventions, decisions, gotchas. The wiki your team never wrote by hand.

L2KNOWLEDGE GRAPH

Decision provenance as a graph. Trace any choice back to its reasons with /hive:why. Opt-in.

L3SEMANTIC INDEX

Optional ChromaDB layer for semantic recall across everything above. Opt-in.

Why Hive

Agents are easy. Organizations are hard.

CROSS-MODEL

Two models, zero bias.

Claude directs and reviews; Codex implements. Cross-model execution cuts cost and eliminates self-review blindness.

TEST SWARM

Tests that hunt.

A dedicated swarm doesn't just run your suite — it writes tests, triages failures, and dispatches the fixes.

LAYERED MEMORY

A hive remembers.

Four memory layers carry architecture decisions, conventions, and hard-won insights across every session and project.

// Open source · Apache 2.0

Built shipping real products. Open to yours.

Hive is Apache 2.0, built at Firefly Events while shipping production software — every workflow in it earned its place on a real deadline. The architecture is deliberately edge-extensible: agents, skills, workflows, and teams all plug in without touching core. If you've ever wanted to shape how AI teams actually work, this is the door.

Join the Hive →
  • Add an agent — a new specialist persona your team is missing.
  • Add a skill — package a workflow you repeat into a reusable command.
  • Add a harness adapter — wire in a new execution substrate and prove the contract.

Put a whole team in your terminal.

One plugin install. Twenty-five specialists. Kickoff to ship.

Get Started