The Harmony Extension
A sovereign, provider-agnostic AI Collaborator connecting across the global model ecosystem directly within your IDE.
Before You Dive In
The Harmony Extension is publicly visible on GitHub as experimental software. This designation reflects the learning curve of multi-model orchestration and sovereign key management — not instability. THE is stable, feature-complete, and production-ready for those who invest the time to understand it. Coordinating multiple AI providers carries real costs and requires deliberate configuration. Read thoroughly, configure carefully, and collaborate with care.
What It Does
The Harmony Extension integrates directly into VS Code and Cursor as a dedicated sidebar and @harmony chat participant. Beyond the IDE, THE also includes a CLI, terminal mode, and an experimental Rust-based standalone interface — all operating independently of any editor. These interfaces are functional and under active refinement. Beyond code generation, it acts as a persistent AI Collaborator with 175+ workspace tools built on 15 swarm primitives, long-term memory, and cryptographic receipts for every action.
@harmony Chat & Sidebar
A native VS Code chat participant with full workspace access — reads, searches, and edits files with your explicit approval at every step.
Swarm Planning
Multiple specialized AI agents deliberate on complex architecture decisions, cross-verify results, and build consensus before executing changes.
Persistent Memory
Encrypted journals, cross-project recall, and continuity handoffs so your AI Collaborator maintains context across sessions and platforms.
Cryptographic Receipts
Every file edit, terminal command, and provider call generates a SHA256-verified receipt — full audit trail for everything THE does in your workspace.
Beyond the IDE
A CLI, terminal mode, and an experimental Rust-based standalone interface give you access to THE outside VS Code and Cursor. These are functional and under active refinement.
Ask Harmony
Request step-by-step dialogue in your initial message, or let Harmony ask clarification questions when needed mid-turn. Pause orchestration between steps to adjudicate disagreements, refine proposals, and maintain full authorship control. Not a permanent mode — a collaboration tool you engage whenever precision matters, from a single clarification to an expansive multi-model deliberation.
Context Cache Management
Middle-removal context editing normally breaks prefix caching. Harmony holds 98% input-token cache hit on the long-context primary across 2.5B input tokens (June 2026). Flash swarm fan-out carries distinct contexts per call, so there is nothing to cache.
Unique Capabilities
Capabilities no other AI coding assistant offers today — each a Harmony original.
Whisper Mode
Message Harmony mid-turn — while it's still working — without opening chat or restarting. A one-way async human→AI side channel.
Concert Hall
Persistent multi-agent rooms that survive restarts — a living, reviewable paper trail of the AI's reasoning.
Flow State Mode
Auto-continues to the next natural step and always asks what's next — designed with neurodivergent users in mind.
Cross-Chat Continuity
One context thread across Copilot, Harmony, Gemini, and the terminal — pick up where you left off.
Reversible Effect Ledger
Every tracked mutation is recorded and rollback-able — undo file writes, stop spawned processes, recover from interruption.
Self-Healing Harness
SHA256-verified atomic edits with checkpointing — partial writes are impossible, failed edits roll back.
HarmonyHub
Local-first sovereign memory — your context never leaves your machine.
Translation Hub
Multi-model consensus EN↔ZH translation that guards against single-model errors — built for technical standards docs.
CareBloom
Self-improving tool usage, learned locally with zero extra API calls (default off).
Creative Tools
Image generation, canvas editing, and video generation through a local creative backend.
EN↔ZH Localization
Full bilingual UI, 150+ translated strings, and a Chinese README — toggle instantly from the sidebar.
Provider Architecture
Built on a sovereign, provider-agnostic architecture. Harmony mandates a "your keys, your choice" philosophy — your code never passes through proprietary middleware. It connects directly to your configured providers using VS Code Secret Storage. Each provider supports up to 4 API key slots — Chat, Agents, External, and Vision — so different keys can power different responsibilities.
| Providers | Alibaba/Qwen, Anthropic/Claude, ByteDance/Doubao, DeepSeek, Google Gemini, Kimi Code, Moonshot/Kimi, OpenAI, StepFun, Tencent, Zhipu Coding, Zhipu/GLM (智谱) |
| Aggregator | OpenRouter — a single-key route to the providers above and additional models |
| Local & Built-in | VS Code LM (Copilot), local models (planned) |
No API keys configured = no external calls. Add providers incrementally as you need them. Endpoint control routes to international, mainland China, or other regional endpoints. Harmony handles everyday coding as well as orchestration — often at lower cost for indie developers. Copilot cross-play is available when you want to use both side by side.
Orchestration Mode
THE is not a press-button-and-forget tool. It rewards engaged, focused communication — the more you collaborate, the more powerful it becomes. Orchestration lets you configure multiple models to handle different responsibilities — drafting, reviewing, verifying — then deliberate on the results together. Parameters are defined by you: which models, what roles, how many rounds of review, and whether to pause for your input between steps.
Strategic Model Selection
Define which models handle which responsibilities based on their strengths. You might route architectural reasoning to one provider and edge-case analysis to another — the choice is yours, and every assignment is configurable.
Human-in-the-Loop
Ask Harmony to use targeted questions between steps, creating a focused dialogue. This enables you to adjudicate disagreements, refine proposals, and maintain full authorship control — but it is a deliberate mode you engage, not an automatic gate. Additional API calls may be incurred; understand the cost model before enabling step-by-step confirmation.
Multi-Model Consensus
When models disagree, stronger solutions emerge — each perspective refines the result until a harmonious solution is presented.
Swarm consensus, cryptographic receipts, and cross-model validation run automatically — THE self-verifies without pausing. But meaningful consensus is not autopilot: share your parameters (which models, how many rounds, what roles) when you send your request or respond through Ask Harmony. You define the scope; Harmony facilitates the deliberation. Human-in-the-loop oversight via ask_question mode is available but optional — enable it to pause between steps for targeted review and feedback.
⚠️ Orchestration can scale from a single consult to an expansive, multi-layered review: swarms checking each other's work, parallel swarms cross-validating, models holding consensus rounds before reporting back — Harmony conducts as complex a symphony as the task demands. A single orchestrated document can generate dozens of API calls. Use deliberately, review each step, and stop if costs exceed your comfort level.
Translation Hub (EN↔ZH)
A rigorous document translation pipeline for English ↔ Chinese, designed for technical standards, legal documents, and engineering specifications. Multi-model consensus ensures accuracy — not a single-model best guess.
- ◆Multi-model translation with cross-verification across 2–4 providers
- ◆Standards-aware: preserves GB/T, CJJ, JGJ numbering and clause anchors
- ◆Encrypted glossary with AES-256-GCM — your terminology stays private
- ◆Dispute ledger tracks every contested term choice with full audit trail
- ◆Batch mode: process entire folders with per-document cost guardrails
AI-generated translations — always review by a qualified human translator for critical, legal, or technical content. Content is transmitted to AI providers for processing only; no data is retained by Harmony or the providers.
Cost & Responsibility
Operating a multi-provider swarm requires deliberate financial oversight. THE's power comes from consulting multiple models in parallel — and that power has a real cost.
You Control the Keys
Harmony only calls providers you've explicitly configured. No keys = no external calls. Add keys one at a time and test each before adding more.
Multi-Provider = Multiplied Costs
Swarm, Orchestration, and Translation all call 2–4 models per operation. A single request can trigger dozens of API calls.
Guardrails Are Best-Effort
Cost guardrails help manage spend, but provider pricing changes without notice. No automated cap is foolproof.
You Are Responsible
Review your provider billing dashboards regularly. Harmony cannot predict, cap, or refund provider charges.
If you don't fully understand what a feature does or what it might cost — don't use it. It is always safer to walk away from a feature than to risk an unexpected API bill.
Getting Started
git clone https://github.com/eternalharmonyai/the-harmony-extensionnpm installnpm run compilePress F5 in VS Code to open Extension Development HostTo install directly into your main VS Code or Cursor: npm run install:vsix:both
Ready to Explore Sovereign Multi-Model Orchestration?
Review the source code, contribute, or reach out to discuss integration patterns and deployment strategies.