🎹 Developer's Words
"I believe playing the piano is also a form of orchestration." The harmony of polyphony — multiple voices — and homophony — a single melodic line. Each voice sings its most beautiful song from its own place, yet when combined, they create one grand, beautiful melody. I believe this structure is no different from AI agents.
Chopin Ballade No.4

Parallel Agent
Orchestration

A mission loop, not a single prompt. Commander, Planner, Worker, and Reviewer agents plan, implement, and verify — and the runtime decides when the work is actually done.

Get Started Quick Start ↓
4
Specialized Agents
Rust
Native Tools
Local
First Memory
Canvas
Knowledge Graph
MIT
Open Source
Node 24+
Runtime
Commander-Led Orchestration

The Commander delegates to Planner, Worker, and Reviewer agents and drives a persisted mission loop. The runtime adjudicates "done" at the idle boundary and, when a task stalls, escalates in order — DECOMPOSE → RE-PLAN → ASK — instead of guessing.

Roles are pluggable: drop a .opencode/agents.json file to swap in your own agent profiles, and retrieval re-weights per role so each agent sees the context it needs.

Local-first mission state under .opencode/ — no external service required.

orchestrator
// task dispatched
Commander decomposing task…
[pool] Sessions A, B, C ready
Planner generating Todo.md
[parallel] Workers A + B spawned
[state] Mission state synced ✓
[gate] Wiring gate: changes verified
Reviewer auditing output…
[verify] TODO + tests passed ✓
COMPLETED
Memory That Forgets Like You Do
A local-first, human-like Ebbinghaus memory — no external vector DB. Notes gain strength on recall and fade when unused; stale ones are de-referenced, never deleted, and every note carries bi-temporal event & ingestion time.
Retrieval by Complementary Senses
BM25 lexical search, tags, and a wiki-link graph are fused with Reciprocal Rank Fusion and biased by role — so each agent recalls the right context the way you'd reach for it.
Verified Completion
The mission loop checks evidence — TODOs, tests, and changed-file wiring — before declaring a task done, then escalates DECOMPOSE → RE-PLAN → ASK if it can't.
Native Tooling
Rust-backed tools (grep, glob, ast, lsp, diff, and more) run in a pooled subprocess for fast, structured results.
Auditable by Design
Every mission leaves a trail — a mission ledger, a working scratchpad, and an Obsidian-compatible knowledge-map canvas you can open and read for yourself.
Safe by Default
Disk writes are opt-in, and sensitive or malicious memories are excluded from prompts — so untrusted notes never steer an agent.
Capabilities at a Glance
Area Capability Detail
Agents Commander-led Planner · Worker · Reviewer delegation
Tooling Native Rust Pooled grep / glob / ast / lsp / diff
Memory Ebbinghaus Notes strengthen on recall, fade when unused — no vector DB
Retrieval Multi-sense + RRF BM25 + tag + wiki-link graph, fused & role-biased
Completion Runtime-verified Evidence gates + self-review + escalation
Installation
Quick Start
01 Install globally
terminal
$ npm install -g opencode-orchestrator
02 Run inside OpenCode
opencode
/task "Implement JWT authentication with audit logs"
📦 npm
opencode-orchestrator
GitHub
agnusdei1207/opencode-orchestrator
📖 Docs
System Architecture