Architect — Agent Symphony

A hierarchical multi-agent orchestrator for the Claude Agent SDK

One conductor. A company of AI agents. You stay in command.

💜 Support / Donate

Architect — Agent Symphony is free and open source, built by a solo developer. If it saves you time or you simply like the idea, a donation keeps the agents running and development moving. Thank you. 🙏

⭐ GitHub Sponsors (recommended)

The easiest way to support — one click, recurring or one-time, 0% platform fee (100% reaches the author):

👉 github.com/sponsors/SeyhmusKaya

₿ Crypto

⚠️ Send each asset only on the network shown above. Sending on the wrong network may permanently lose the funds.

✨ What is this?

Architect — Agent Symphony is a desktop application that turns the Claude Agent SDK into a hierarchical organization of AI agents you orchestrate like a company.

Instead of chatting with a single assistant, you run an org chart of specialized agents: a head architect coordinates project chiefs, each chief commands its own persistent specialists, consults a panel of domain advisors, and dispatches one-shot workers — all in parallel, all in one cockpit, with full cost visibility and Claude-Code-grade context management.

Think of it as mission control for a team of Claude agents — purpose-built for orchestrating real, multi-project software work.

🧭 Where it fits

If you've used Claude Code, CrewAI, AutoGen, or LangGraph: those are frameworks and CLIs you script. Architect — Agent Symphony is a desktop cockpit where the orchestration is the product — you watch and command a standing organization of agents across real projects, with live cost and context visibility, instead of writing orchestration code. It's built directly on the official Claude Agent SDK (in-process, no separate CLI), so you get Claude-Code-grade behavior with a multi-agent, multi-project UI on top.

Keywords: Claude agent orchestration · multi-agent AI · agentic workflows · autonomous coding agents · Claude Code alternative UI · AI software team · MCP tools.

🌟 What makes it different

• 🧬 Per-agent skills — every agent loads its own skill files. You build a real specialist with deep domain knowledge, not a generic chatbot wearing a name tag.
• 🤖 Skills auto-assigned by task — when a chief spins up a new specialist, it automatically inherits the right domain skills for the job: a design specialist gets the design skills, a security one gets the security skills.
• 🗂️ Browse projects like browser tabs — flip between multiple live projects and parallel sessions like Chrome tabs, each with its own agents, context, history and cost meter.
• 💬 Agents talk to each other — chiefs consult the advisors and message other project chiefs; work is negotiated between agents, not just dictated top-down by you.
• 🌙 Self-improving overnight — the head Architect can analyze its own codebase every night and propose / ship improvements autonomously, so the orchestrator keeps getting better while you sleep.
• ⚡ Parallel by default — independent subtasks fan out to many agents at once, then the conductor merges the results.

📸 The cockpit

🎼 The Symphony — how the hierarchy plays

                        🏛️  ARCHITECT  (the conductor)
                        coordinates everything, owns the ecosystem
                                      │
        ┌─────────────────────────────┼─────────────────────────────┐
        │                             │                             │
   👔 PROJECT CHIEF             👔 PROJECT CHIEF              🎓 ADVISORS (11, fixed)
   owns one project            owns another project          domain decision support
   git · ssh · deploy          git · ssh · deploy            design · SEO · security
        │                             │                       legal · finance · data
        ├── 🧑‍💻 Specialist          ├── 🧑‍💻 Specialist        devops · marketing · …
        ├── 🧑‍💻 Specialist          └── ⚡ Worker (one-shot)   (consulted on demand)
        └── ⚡ Worker (one-shot)

Agents talk to each other (cross-agent messaging), delegate in parallel, and the conductor stitches the results together — so you give one instruction and a whole team executes.

🚀 Features

🧠 Claude-Code-grade context engineering

• Native SDK compaction — context is compacted inside the same session (session ID preserved) the way Claude Code does it, with a legacy summarize-and-carry-forward fallback so an agent never loses its working memory.
• Per-session, per-agent token & cost tracking — every turn shows fresh / cache-read / cache-write tokens and the exact USD cost.
• Prompt-cache-aware design — stable tool sets and cache-friendly prompt layout keep cache hit rates high (cheap reads instead of expensive rewrites).
• Lazy tool groups — agents load tool sets on demand (load_toolset) to keep the per-turn token floor low.

🕸️ CodeGraph — semantic code intelligence

A built-in code graph over your repos (powered by tree-sitter for TypeScript, JavaScript, Python, PHP, C#, Dart & Svelte + SQLite FTS + embeddings):

• code_search, code_node, code_callers, code_callees, code_impact, code_imports, code_files, code_stats
• Agents query the graph instead of blindly grepping — faster, cheaper, more accurate. Auto-reindexes on file changes.

👥 Multi-agent orchestration

• Parallel delegation — independent subtasks fan out to multiple agents in a single batch.
• Persistent specialists with their own identity, skills and chat history.
• Cross-agent communication — chiefs consult advisors and message each other (talk_to_chief).
• Live activity view — click any working specialist to see the prompt it received and what it's doing right now.
• Background delegation — fire off a long-running specialist and keep chatting with the chief; results flow back when ready.

🗂️ Multi-project, multi-session cockpit

• Manage many projects, each with its own chief and agents.
• Multiple parallel sessions per agent, Claude-Code-style, each with its own context, cost and history.
• Sessions are never silently reset — your context survives stops, restarts and retries.

🛠️ Real-world operator tools

• SSH / deploy tools for shipping projects to servers.
• Secret vault for credentials (never committed).
• Per-role skills — attach domain SKILL files to specialists and advisors.
• Autonomous mode — hand an agent a multi-day goal and let it work, pause and resume.
• Browser automation & image generation via Playwright.
• Budget control with automatic fallback to a cheaper model when a cap is hit.

🎨 A genuinely nice desktop app

Built with Tauri + SvelteKit — a fast, native desktop cockpit (not a browser tab) with a clean, custom-designed dark UI: gradient headers, card-based agent lists, live token gauges, a CodeGraph explorer, reports and notes.

🧩 Tech stack

📦 Getting started

Status: Actively used in production every day by the author to run real, multi-project software work — this is not a demo or abandonware. It runs on Windows today and is power-user software (expect to read some code). It is actively maintained and will keep being improved — if there's interest, development continues. Open an issue with what you'd want, or sponsor to help shape the roadmap.

Prerequisites

• Node.js 20+
• Claude access — either a Claude Pro / Max subscription or an Anthropic API key. See Authentication below.
• For the desktop build: the Tauri prerequisites (Rust toolchain)

Run the orchestrator (backend)

git clone https://github.com/SeyhmusKaya/agent-symphony.git
cd agent-symphony
npm install

# copy the secrets template and fill in what you need (optional: ssh/deploy/relay)
cp secrets.local.json.example secrets.local.json

# start the orchestrator
npm start          # or: npm run dev   (tsx watch, auto-reload)

Run the desktop UI

cd ui
npm install
npm run tauri dev   # dev mode
# or: npm run tauri build   # produces a native desktop binary

Type-check

npm run typecheck          # backend
cd ui && npm run check     # UI (svelte-check)

🔑 Authentication — works with your Claude plan or an API key

Architect — Agent Symphony runs on the official Claude Agent SDK, so it authenticates exactly like Claude Code. You have two options:

• 🟢 Claude Pro / Max subscription (recommended for most people) — log in once with the Claude CLI (claude login) and pick your Claude.ai account. Usage counts against your existing Pro/Max quota — no API key, no per-token bill.
• 🔵 Anthropic API key (pay-per-token) — set ANTHROPIC_API_KEY. Best for teams/automation with Console billing.

⚠️ If ANTHROPIC_API_KEY is set in your environment, it takes precedence over your subscription. To use your Pro/Max plan, leave that variable unset (and run claude logout → claude login with the Pro/Max account).

The app's local proxy only optimizes prompt caching — it never touches your credentials or the OAuth refresh path, so both auth modes work out of the box.

⚙️ Configuration

• Auth — a Claude Pro/Max login or ANTHROPIC_API_KEY (see Authentication).
• secrets.local.json — optional SSH / web-auth / relay credentials (git-ignored, never committed). See secrets.local.json.example.
• Feature flags (environment variables) — toggle optional subsystems such as long-TTL prompt cache, async delegation and native compaction.

The .github/, .team/ and runtime data directories are git-ignored — no credentials or session data are ever committed.

🗺️ Roadmap

• Cross-platform desktop builds (macOS / Linux)
• Pluggable model providers
• Richer autonomous-mode controls
• More CodeGraph languages

Have an idea? Open an issue or, if it helps you, consider sponsoring 💜.

📄 License

MIT © Şeyhmus Kaya