Event Horizon — Multi-agent orchestration for AI coding agents

Turn any AI coding agent into a project manager. One command creates a plan, spawns a team of AI agents, assigns roles, and manages the entire project -- while you watch it happen live in a cosmic visualization.

Event Horizon Demo

Get started in 30 seconds

Click the rocket icon (top-right of any editor tab) or press Ctrl+Shift+E H
Click Connect and choose your agent
Start coding -- your agent appears as a planet

No accounts. No config files. No API keys. Everything runs on your machine.

No agent running? Click Demo to see the universe in action with simulated agents.

What happens when you use Event Horizon

Before: you manage every agent manually

You open three terminals. Claude Code builds the API. OpenCode writes tests. Copilot updates docs. They all edit the same project. Claude overwrites OpenCode's changes to server.ts. Nobody notices. The build breaks. You spend 20 minutes untangling.

After: your agent manages the team for you

/eh:create-plan Build a REST API with auth, database layer, and tests

Your agent becomes the orchestrator. It analyzes the work, breaks it into parallel tasks, spawns worker agents in visible terminals, assigns roles, enforces file locks, tracks the budget, retries failures, and reports back when everything is done. You watch it happen on a live Kanban board and a cosmic visualization where every agent is a planet.

Works with every major AI coding agent

Agent	Hooks	MCP Tools	File Locking	Spawnable	Token Tracking
Claude Code	Yes	Yes	Yes	Yes	Yes
OpenCode	Yes	Yes	Yes	Yes	Yes
GitHub Copilot	Yes	Yes	--	--	Partial
Cursor	Yes	Yes	--	Yes	Yes

One-click connect. MCP server auto-registered. Hooks auto-updated on every activation. Mix and match agents freely -- Claude orchestrating OpenCode workers is a first-class workflow.

Core capabilities

Multi-agent orchestration

The agent that creates a plan auto-becomes the orchestrator with elevated MCP tools to spawn agents, assign tasks, monitor the team, and control budgets. Spawned agents run in visible VS Code terminals -- full transparency, click any planet to focus its terminal.

Plan coordination

Atomic task claiming (no duplicate work), dependency resolution (blocked tasks auto-unblock), cascade failure with auto-retry and model escalation (haiku fails, sonnet retries, opus escalates). Smart task recommendations score agents by role match, historical performance, and current load.

File locking and worktree isolation

When Agent A edits a file, Agent B is hard-blocked -- the tool call doesn't execute. Locks refresh on writes and release on agent termination. For full isolation, agents can work in their own git worktrees with auto-merge on completion.

Shared knowledge

A live knowledge base where humans and agents contribute context in real-time. Workspace knowledge persists across sessions. Plan knowledge lives with the plan. Auto-discovers CLAUDE.md, .cursorrules, copilot-instructions.md, and other instruction files.

Budget and cost controls

Per-plan spending limits with warning at 80% and hard stop at 100%. Tiered model selection tries the cheapest model first per task complexity. Failed verification auto-escalates (haiku to sonnet to opus). Cost Insights panel shows cache efficiency, duplicate reads, and actionable recommendations.

Agent roles and profiling

Eight built-in roles (researcher, planner, implementer, reviewer, tester, debugger, orchestrator, context-optimizer) plus custom roles. The profiler tracks success rate, speed, and cost per agent type per role -- so you know which agent is best at what.

Live visualization

Every agent is a planet. The cosmic metaphor encodes real information:

Visual	Meaning
Planet type (gas, rocky, icy, volcanic)	Agent type (Claude, OpenCode, Copilot, Cursor)
Golden star with emission rays	Orchestrator managing the plan
Pulsing ring	Agent is thinking
Amber breathing ring	Waiting for your input
Red glow	Error state
Orbiting moons	Active subagents
Ships between planets	Data transfers between cooperating agents
Lightning arcs	File collision -- two agents editing the same file
Asteroid belt	Workspace group (agents sharing a directory)
Orbital debris	Plan tasks (shape and color encode status)

Operations dashboard

Full-screen dashboard (Ctrl+Shift+E O) with agents/plans sidebar, metrics overview, file activity heatmap, searchable event logs, timeline swimlanes, Kanban board with dependency DAG, role assignments with performance profiles, cost insights, and a shared knowledge panel.

40+ MCP tools

All agents access coordination tools via the MCP server (auto-registered on connect):

eh_load_plan eh_get_plan eh_list_plans eh_claim_task eh_update_task eh_retry_task eh_recommend_task eh_verify_task eh_archive_plan eh_delete_plan eh_send_message eh_get_messages eh_check_lock eh_acquire_lock eh_release_lock eh_wait_for_unlock eh_list_agents eh_file_activity eh_write_shared eh_read_shared eh_get_shared_summary eh_delete_shared eh_list_roles eh_assign_role eh_get_agent_profile eh_recommend_agent eh_heartbeat eh_get_budget eh_get_session eh_get_cost_insights eh_spawn_agent eh_stop_agent eh_reassign_task eh_get_team_status eh_auto_assign eh_create_worktree eh_remove_worktree eh_request_budget_increase eh_sync_skills eh_search_events

Nine bundled skills handle common workflows so agents don't need to memorize tool names:

Skill	What it does
`/eh:create-plan`	Generate a plan with parallel tracks, dependencies, acceptance criteria, and verify steps
`/eh:work-on-plan`	Claim tasks, implement, self-verify against acceptance criteria, mark progress
`/eh:orchestrate`	Manage a plan as orchestrator -- spawn workers, assign tasks, monitor, handle failures
`/eh:verify-task`	Batch-verify completed tasks by running their verify commands
`/eh:optimize-context`	Reduce token costs by analyzing and splitting instruction files
`/eh:plan-status`	View progress, blocked tasks, active agents, available work
`/eh:research`	Explore codebase and output structured findings
`/eh:review`	Code review with severity levels and verification pipeline
`/eh:test`	Write tests following project conventions
`/eh:debug`	Diagnose bugs, trace root cause, apply minimal fix

28 achievements

Milestones that track your multi-agent journey -- from spawning your first agent to surviving file collisions, catching UFOs, and diving astronauts into black holes. Some are secret. Some have tiers. All persist across sessions.

Privacy

100% local -- server on 127.0.0.1:28765, nothing leaves your machine
Zero agent overhead -- if Event Horizon is closed, agents run identically
No telemetry -- no analytics, no tracking, no data collection

Links

MIT License with Commons Clause -- see LICENSE.

Event Horizon

HeytalePazguato