Skip to content
/ hive Public

Hive: A Lightweight Multi-Agent Coding Orchestrator

Notifications You must be signed in to change notification settings

tnguyen21/hive

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

364 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Hive: Multi-Agent Coding Orchestrator

Alpha software. Expect rough edges.

Hive coordinates parallel coding workers against a shared issue queue, then helps you review and finalize work.

When To Use Hive vs Claude Code

Task Just use Claude Code Use Hive
Small bug fix in one area Yes -
Single focused feature Yes -
Feature + tests + docs + migration - Yes
Refactor across multiple modules - Yes
Spec that naturally splits into subtasks - Yes

Rule of thumb: if you'd split the work into separate tasks anyway, Hive is the better fit.

How It Works

Hive has three moving parts:

Issues are your task board, stored in SQLite. Each issue is a unit of work: a bug fix, a feature, a refactor step. You create them, and the system tracks their lifecycle from open through in_progress, done, and finalized.

Workers are Claude Code sessions that pick up issues and implement them. Each worker gets its own git worktree, a full isolated copy of the repo. Workers run in parallel, up to your configured concurrency limit. When a worker finishes, its branch goes through merge validation (rebase, tests, optional refinery review) before anything touches main.

The Queen is how you drive Hive. It's a Claude Code session that acts as your project manager: it reads your spec, explores the codebase, decomposes work into issues, monitors progress, and handles failures. The Queen doesn't write code. It plans and coordinates, and you approve the plan before any issues are created.

The typical flow:

  1. You describe what you want to the Queen
  2. The Queen explores the codebase, proposes a plan, and waits for your approval
  3. On approval, the Queen creates issues (with dependencies if needed)
  4. The daemon picks up issues and spawns workers in parallel
  5. Workers implement, the merge pipeline validates
  6. You review completed work and finalize what lands on main

What It Looks Like

Hive demo: Queen session and live status

Left pane: The Queen session in Claude Code. Here it's exploring the codebase, reading project files, and responding to an escalated issue that a worker couldn't resolve on its own.

Right pane: watch -n 1 hive status giving a live dashboard — issue counts, active workers, merge queue throughput, and any items that need your attention. In this snapshot, 74 issues have been finalized, 115 branches merged, and one escalation is flagged for the operator.

This is the typical working setup: you interact with the Queen on the left while the right pane gives you a constant read on system state.

Monitoring with hive status

hive status is the single most useful command for keeping tabs on a run. It shows:

  • Issue breakdown: how many are open, in progress, done, finalized, escalated, or cancelled
  • Active workers: how many are running vs your configured concurrency limit
  • Merge queue: pending merges, total merged, total failed
  • Needs attention: escalated issues or failures that require your intervention
  • Daemon state: whether the daemon is running and where its log lives

Pair it with watch for a live dashboard:

watch -n 1 hive status

This lets you monitor progress without interrupting your Queen session. When something needs attention — an escalation, a stuck worker, a merge failure — you'll see it immediately.

Quick Start

Install

If you want hive available across all repos:

cd hive
uv tool install -e .
uv tool update-shell   # ensures uv's tool bin dir is on PATH

Or install into a project venv:

cd hive
uv venv --python 3.12
source .venv/bin/activate
uv pip install -e ".[dev]"

Notes:

  • Use uv tool install -e ".[dev]" if you want the dev extras too.
  • If you previously ran uv tool install ., you may need --force to refresh: uv tool install --force .

First Run

cd your-project

# 1. One-time setup, creates .hive.toml with your project config
hive setup

# 2. Seed project conventions so workers know your norms
hive note "Always run pytest before committing" --category context
hive note "Use ruff for linting, line-length 144" --category context

# 3. Launch the Queen, your main interface
hive queen

The Queen auto-starts the daemon if it isn't running. From here, describe what you want built. The Queen will explore the code, propose a decomposition, and wait for your go-ahead before creating issues and dispatching workers.

If you prefer to manage issues manually instead of using the Queen:

hive create "Add user auth" "Implement JWT login flow"
hive create "Add auth tests" "Unit tests for the JWT flow" --depends-on <auth-issue-id>
hive start       # foreground with live status
hive start -d    # or detached as a background daemon

Reviewing and Finalizing Work

By default, Hive runs in manual review mode. Completed work sits in done status until you explicitly finalize it. Nothing merges to main without your approval.

# See what's ready for review
hive review

# Inspect a branch (review command prints these hints per issue)
git diff main...agent/<agent-id>

# Finalize: merges to main and cleans up the worktree
hive finalize <issue-id> --resolution "reviewed, looks good"

If you enable merge_queue_enabled in your config, Hive will auto-merge branches that pass rebase + tests. A refinery LLM session handles conflict resolution and integration checks. Issues that the refinery can't resolve cleanly get escalated for your attention.

Teaching Workers Your Conventions

Notes are persistent context that every worker receives in its prompt. Use them to encode project norms:

hive note "Run ruff check and pytest before committing"
hive note "All API endpoints need OpenAPI docstrings"
hive note "Use factory pattern for test fixtures" --category pattern

Categories are optional. The default is context; use --category pattern, --category gotcha, or --category dependency for more specific classification.

Workers also write notes during execution (discoveries, gotchas, dependency observations) which get relayed to sibling workers on the same epic. This is how parallel workers stay loosely coordinated without sharing a context window.

Cost and Performance

Hive spawns multiple Claude Code sessions. Costs scale linearly with concurrency and task complexity.

For your first run, set max_agents = 3 to keep concurrency low. Validate that the review loop works for your project, then scale up once you're comfortable.

Hive tracks token usage per issue and per run. Use hive metrics --costs for rough USD estimates, and configure max_tokens_per_issue / max_tokens_per_run as guardrails.

Essential Commands

Command What it does
hive queen Launch the Queen (main interface)
hive setup One-time project config
hive create Create an issue manually
hive start / hive stop Start/stop the daemon
hive status System overview: issues, workers, merge queue
hive list List issues (filter with --status)
hive show <id> Issue details, deps, recent events
hive review Review completed work before finalizing
hive finalize <id> Finalize and merge to main
hive retry <id> Retry a failed issue
hive note Add persistent context for workers
hive doctor Health checks (use --fix to auto-repair)
hive debug Full diagnostic dump for bug reports

See hive -h for the full list, including logs, agents, merges, metrics, epic, and dep.

Configuration

Config is layered (later overrides earlier):

  1. Built-in defaults
  2. ~/.hive/config.toml (global)
  3. .hive.toml (per-project)
  4. Environment variables

Key settings:

Setting Default What it controls
HIVE_BACKEND claude Backend: claude or opencode
HIVE_MAX_AGENTS 10 Max concurrent workers
HIVE_TEST_COMMAND Test command run during merge validation
HIVE_MERGE_QUEUE_ENABLED true Auto-merge vs manual review mode
HIVE_WORKER_MODEL claude-sonnet-4-5-20250929 Model for workers
HIVE_REFINERY_MODEL claude-opus-4-6 Model for merge refinery
HIVE_MAX_TOKENS_PER_ISSUE 200000 Per-issue token budget
HIVE_MAX_TOKENS_PER_RUN 2000000 Per-run token budget

Architecture and Internals

For orchestration design, schema details, merge pipeline internals, and backend abstraction, see docs/TECHNICAL_DESIGN_DOC.md.

Bug Reports and Feedback

Please send me bug reports, feature requests, and other cool ideas!

Run hive debug and include the output in your report. It collects system info, config, daemon state, doctor checks, and recent logs into a single pasteable bundle:

hive debug          # human-readable
hive debug --json   # machine-readable

Development

# tests
uv run pytest

# lint + format
uvx ruff check src/ tests/
uvx ruff format --line-length 144 src/ tests/

About

Hive: A Lightweight Multi-Agent Coding Orchestrator

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages