Synergy – Persistent AI‑Agent Workspaces

What it is

• An open‑source, local‑first platform that lets AI agents (e.g., Atlas, Forge, Pico) keep their state—history, files, browser sessions, and tool configurations—across multiple interactions and devices.
• Provides a single runtime that can be accessed via a web workbench, an Electron desktop app, a CLI, or a server‑API/SDK, so the same session is available everywhere.

Key capabilities

Feature	Why it matters
Durable sessions	Work that outlives a single chat is saved in a “Scope” (project or home) with full history, even when the LLM context is trimmed.
Agent coordination	Primary agents can delegate to specialist sub‑agents, run reusable Blueprints, and orchestrate complex workflows (Light Loop, Boss Mode).
Files & Browser in context	Files can be edited, moved, or deleted while a web page stays open in an integrated Chromium browser; website logins are persisted per session.
On‑demand compute	Execution can run in native processes or Docker containers; API‑only tasks use no container, and storage can live on local disk, S3, or OSS.
Knowledge compounds	A Library stores reusable memory, notes, and evaluated experiences with reward‑signal analytics.
Visual results	Agents can stream interactive charts, simulations, or HTML that remain attached to the conversation for later review.
Extensible & local‑first	Users can add providers, tools, Skills, plugins, and custom channels while keeping all data under their own control.

Agents

• Atlas – general‑purpose assistant.
• Forge – coding‑focused agent (used in benchmark).
• Pico – lightweight, quick‑turn tasks.

Benchmarks

• On the DeepSWE v1.1 coding benchmark (113 repo‑engineering tasks) the Forge agent with Synergy raised Pass@1 from 53 % to 67.3 % (+14.3 pp) at $0.54 per task, placing it on the cost‑performance Pareto front.
• Reported average cost per passed task: $0.60 for ~2.3 h of compute, 7–40× cheaper than top‑5 official models.

How to get started

1. Desktop – download the installer (.pkg, .exe, or .deb) from the latest GitHub release; it also installs the synergy CLI.
2. CLI / Web – run the one‑line installer:

   curl -fsSL https://raw.githubusercontent.com/SII-Holos/synergy/main/install | bash
   

   Then configure a model provider (synergy config wizard), start the runtime (synergy start), and open the web workbench (synergy web).
3. Run a task – e.g., synergy send "summarize this repository".

Runtime commands – status, logs, doctor, stop, upgrade, uninstall (with fine‑grained flags to keep data or only remove a specific installation channel).

Development

• Monorepo built with Bun and TypeScript ESM.
• Core packages: agent-runtime, harness, local-runtime, cli, and presets for composable capabilities.
• Rust/Cargo needed for native PTY and Linux process‑ownership library.
• Common dev scripts: bun dev server, bun dev app --open, bun dev desktop, bun dev send "msg".
• Tests are run per package (bun test) and aggregated via bun run quality.
• Plugins can be created without cloning the repo using the synergy-plugin-kit.

Product surfaces

Surface	Use case
Web	Full workbench (sessions, notes, library, agenda, plugins).
Desktop	Electron app with native browser, file picker, and OS‑level computer‑use permissions.
CLI	Runtime management, one‑off send commands, diagnostics.
Server API / SDK	Programmatic integration for other tools.

Community & governance

• Maintained by the Holos team at the Shanghai Innovation Institute.
• Open‑source under the MIT license; contributions welcomed via the usual pull‑request flow.
• Detailed contribution, security, and code‑of‑conduct guidelines are provided.

---

Bottom line: Synergy is a full‑stack, locally hosted AI‑agent platform that treats an agent’s work as a persistent, recoverable workspace, enabling long‑running, multi‑tool projects to be managed from any interface while keeping data under the user’s control.