AI OrchestraAI-ORCHESTRA WEB CHAT
Menu
Sign inGet the extension
We Sell No Models · 100% BYOK · The Power Aggregator

You chair. The AIs argue.

AI Orchestra sells no AI models. Users bring their favorite Frontier models with their own API keys — Anthropic, OpenAI, Google, xAI — or any model with an API, including cloud resellers (OpenRouter, Groq, Together, OpenCode), local Ollama, or self-hosted LLMs. Create a workspace with the models of your choice and the providers of your choice. Or just one for all roles. A tool optimized to create the ultimate custom AI team with real-time Context and logic.

Frontier APIs · Cloud Resellers · Local & Self-Hosted LLMs. Direct connection — no proxy, no token markup.

AI Orchestra · illustrative session
task: planmode: Raise-Handperm: ask
  1. 01 · Drafter · Anthropic
    Proposed plan: add the parser, then ship.
  2. 02 · Reviewer · OpenAI · read-only
    BLOCKER — nested code fences have no test. Pause before freezing this plan.
  3. 03 · You · meeting chair
    Admitted ✓ Add that test to the acceptance criteria.
  4. 04 · Drafter updates PLAN.md
    Ship after implementation.
    + Nested-fence regression test must pass before release.
A read-only seat changes the plan through your decision. See a real session →
The Power Aggregator

We sell no AI models. You bring the brains. We build the ultimate team.

No model vendor will ever build cross-vendor orchestration, because each wants you trapped inside its walled garden. AI Orchestra sells zero inference tokens. Instead, we are the power aggregator that optimizes Frontier APIs, wholesale cloud resellers, and private self-hosted models into a disciplined, governed engineering unit.

1. Frontier Model APIs

Direct BYOK connection to Anthropic (Claude 3.7 / 3.5), OpenAI (GPT-4.5 / o3 / o1), Google (Gemini 2.5 Pro), and xAI (Grok 3). Direct vendor calls on your own developer keys — full reasoning depth, zero token markup.

2. Cloud Resellers & Gateways

Tap wholesale compute on OpenRouter, Groq (500+ tok/s LPU speed), Together AI, DeepInfra, Fireworks, SiliconFlow, CheaperInference, or OpenCode for open-weight powerhouses like DeepSeek R1/V3, Qwen 2.5 Coder, and Llama 3.3.

3. Local & Self-Hosted LLMs

Total privacy and offline sovereignty with Ollama, vLLM, LiteLLM, or your company’s internal VPC inference endpoints. Zero token fees, zero code leaves your environment.

Or run a combination of all three in the same room

Seat a Drafter on Claude 3.7, an independent Reviewer on DeepSeek R1 via Together AI, an Indexer on Groq LPU for instant repository mapping, and a local Ollama model for private security checks. Every seat works the same repository, disagrees out loud via hand-raises, and logs every decision to an audit trail in your repo.

Your work has a home

One Project.
Wherever you work.

A goal, a constitution, your AI roles and one workspace. Keep a large codebase local, or follow the path toward cloud work without a local folder.

Available in the extension

Keep it local

Your repository, your machine, your development tools. Keep your files local, with room for a large codebase. Model requests still send your prompts and selected context directly to the providers you use.

Keep the goal, role contracts and work together. A loaded Project currently writes source inside its workspace folder; verify the target before editing an existing repository.

Source of truth
Your declared local Project folder
LOCAL PROJECT

Your machine

  • 01One goalWhat the team is here to accomplish
  • 02Roles & constitutionWho does what, within your rules
  • 03Workspace & decisionsThe work and the reasons behind it
Stop copy-pasting your AIs

You are already the orchestra leader. You just do it by hand.

Most serious developers already pay for three or four model subscriptions and run them side by side, serving as the copy-and-paste layer between them. AI Orchestra is built for you. Every model sees the same context, disagreements reach you instead of being averaged away, and the decision trail outlasts the session.

By hand, in five tabs
  • ·You paste the same context into every model, every time.
  • ·Two answers disagree. You reconcile them in your head, silently.
  • ·Nothing objects to the writer — you are the only reviewer in the loop.
  • ·The plan gets written afterwards from memory, if at all.
  • ·Six weeks later, nobody can say why the decision went that way.
Chaired, in one room
  • One repo. Every model. Context is shared once across all seats.
  • Drafter writes. Reviewer objects. You see the disagreement before code ships.
  • Models raise hands at BLOCKER, CONCERN, or FYI. You call it.
  • PLAN.md and EXECUTION.md fall out of the conversation, not after it.
  • Every routing call and hand-raise is logged to an audit trail in your own repo.

No proxy. No token markup.

Bring your own API keys. Requests go direct to Anthropic, OpenAI, Google, xAI, cloud resellers, or your local Ollama. We are paid for coordination, never for tokens.

Your keys never leave you

Keys stay on your machine or encrypted in your private Vault. Nothing is proxied, logged, or resold. You keep complete sovereignty.

Audit trail in your repo

Every turn, vote, and hand-raise is recorded in your repository or local folder. The record of who decided what outlasts the session.

Test it once, on something real

Pick a ~50-line feature you actually want. Choose a Drafter and an independent Reviewer within your seat capacity. Watch for one thing: did a read-only seat raise something the Drafter missed — and did admitting it change the plan? If that never fires, this is the wrong tool for you, and we would rather you found out in an afternoon.

How it works

Three rules turn a pile of models into a working room.

Give each AI a real role

Drafter writes to disk. Reviewer objects. Indexer holds the repo map. Each seat is a declared role with its own model and its own permissions — not another copy-paste chatbot.

They raise hands. You call it.

An agent can’t interrupt — it raises a hand at BLOCKER, CONCERN or FYI and waits. You see who is objecting to what before code ships, then admit or dismiss each one. Silence is a decision too.

Every decision, on the record

Every seat works from shared project context. The writer updates the plan; read-only seats challenge it through hand-raises. Every routing call and hand-raise is logged to an audit trail in your repo.

The seats & roles

A seat is a job, not a subscription to another chatbot.

Your plan sets how many agents can hold a seat at once. Roles are yours to assign — bring your own provider keys, seat multiple specialized engineers on one codebase, or assign independent review, indexing, and verification mandates across your frontier models.

Architect
System Design · Coordination
Leads system design and room coordination. Establishes module boundaries, evaluates trade-offs, coordinates the table, and outlines the technical plan before any file is touched.
Drafter / Multiple Engineers
Writes to Disk · Line-Locked
Writes to disk. The only seat that can change your files. Assign multiple specialist seats — Frontend Engineer, Backend / API Engineer, Database Architect — with revision-locked line edits and strict WriteGate confirmation.
Reviewer
Diff Inspector · Zero Edits
Reads the diff, raises BLOCKERs before the commit, never edits. Inspects proposed changes, checks architectural guardrails, and halts execution on regressions or safety violations.
Indexer
Repo Map · AST Symbols
Keeps the repo map and the knowledge layer current for everyone. Indexes symbol hierarchies, tracks dependency graphs, and flags cross-module blast radiuses across the workspace.
Researcher / Analyst
Docs & Live Web Intel
Investigates external documentation, live APIs, and third-party dependencies. Gathers grounded empirical evidence, benchmarks, and API specs without touching or polluting source code.
Cold Reader
Independent Verification
Eliminates consensus and confirmation bias. Recomputes math, algorithms, and logic independently from raw input and requirements without seeing teammate drafts or intermediate answers.
Red Team
Adversarial Stress Testing
Challenges the plan deliberately. Assign the role within your available seats. Actively hunts for security vulnerabilities, race conditions, malicious edge cases, and unexpected failure modes.
Synthesizer
Decisions · PLAN.md
Collapses a long exchange into the decision that got made. Distills multi-seat debates into actionable execution, updates PLAN.md, and prepares the EXECUTION.md manifest.
Strategist
Roadmap & Milestones
Anchors engineering to product objectives. Evaluates milestone sequencing, commercial scope, technical debt, and business trade-offs before resources are spent.
Pricing

Priced by the seat, because that’s what you’re adding.

Higher account tiers provide more seats and more Vault storage. Local folders remain yours; cloud Project allowances vary by tier.

Basic
Free forever
2 agent seats
5 MB Vault storage allowance
Basic Plus
$14 /mo
3 agent seats
50 MB Vault storage allowance
Featured
Pro
$20 /mo
4 agent seats
250 MB Vault storage allowance
Pro Plus
$24 /mo
5 agent seats
1 GB Vault storage allowance
Max
$30 /mo
6 agent seats
5 GB Vault storage allowance

Two seats can collaborate. Add a third when you want a tiebreaker or another specialist. Compare monthly and annual billing →

Why this room?

Why not just use Cursor or Copilot?

No model vendor will build cross-vendor orchestration, because each wants users inside its own ecosystem. Only a company that sells no model can seat competing models at the same table. AI Orchestra focuses on a specific collaboration pattern: different vendors occupying declared roles in one shared conversation, with objections you admit or dismiss.

Across companies

Seat Claude, GPT, Gemini, Grok, cloud resellers, and Ollama together using your own keys. Each seat has a defined job in the same discussion.

The AI reviewer that objects

Drafter writes. Reviewer objects. A read-only reviewer raises a BLOCKER before code is committed, so you see disagreement before code ships.

Audit trail in your repo

Every routing call, hand-raise, and vote is recorded in your repository or local folder. The record of who decided what outlasts the session.

Model choice alone is not the distinction. See the current Cursor model documentation and Copilot model documentation. Copilot is not a seatable provider in AI Orchestra.

Read the real journey evidence →
Your keys never leave you

Calls go straight to your provider. Nothing proxied.

Your keys stay local; model requests go directly to your providersProvider keys are held in VS Code SecretStorage. The extension sends requests directly to Anthropic, OpenAI, Google or xAI. A separate connection handles Supabase account and entitlement data. Optional audit sync is planned, not enabled by default. Supabase is not in the model request path.Your VS Code extensionKeys: local SecretStoragePrompts + selected contextYour model providersAnthropic · OpenAIGoogle · xAISupabase · separate account connectionAccount + entitlement dataOpt-in audit: planned · no provider keys or model trafficDirectSign-in + plan access
No AI Orchestra inference proxy. Provider usage is billed by the provider. Local chat stays local by default. Vault storage for Projects and other work is a separate, opt-in roadmap feature.
Read the privacy details →
Chat Web App · foundation preview

Co-Work — the same room, in the browser

The extension is for coding. Co-Work will bring the shared AI team to chat-shaped work, with the same human chair, roles, and hand-raise discipline.

Legal review

Have one role analyze a document and another challenge its reasoning, with a professional reviewing the result.

Research synthesis

Compare sources, surface conflicting evidence, and turn the discussion into a traceable synthesis.

Compliance audit

Assign review responsibilities, surface missing evidence, and keep the human decision attached to each concern.

The first signed-in Chat workspace now lives here. You can configure a team and draft a first message without creating a Project. Today that draft stays in your browser tab: providers, execution and Vault saving are still being connected.

Start with no Project

Sign in to explore the honest foundation while the live provider and Vault path is built. The page labels every unavailable capability in place.

Open the Chat preview