Backchat

A multi-agent desktop for Agent Client Protocol (ACP) harnesses, tools, and outputs.

Visit the official website →

New: DeepSeek Harness support through dsh-plugin and dsh-acp.
_{Use your existing dsh profile or the standalone ACP server from the same desktop.}

Backchat gives developer agents a shared desktop without taking away the things that make each harness useful. Bring multiple agents into one conversation and keep the built-in browser, files, terminal, MCP Apps, and Codex plugins close.

Why Backchat

• One workspace, many agents. Run Claude Code, Codex CLI, DeepSeek Harness, Gemini CLI, OpenCode, Hermes, or OpenClaw from the same desktop surface.
• Agent-aware, not vendor-shaped. Backchat surfaces the capabilities, commands, models, modes, and permissions reported by the active harness instead of inventing a generic capability layer.
• Project-scoped context. Sessions are grouped by their working directory, so files, terminals, browser state, and conversations stay attached to the project where the work happened.
• Projects can span related folders. One primary folder supplies the chat cwd, Git context, and project instructions; secondary folders extend the agent's file scope through ACP additionalDirectories.
• Built-in tools that stay in the flow. Browse the web, inspect files, start a terminal, attach images, and work without leaving the conversation.
• Native extensions and apps. Render interactive MCP Apps directly in a turn and use Codex plugins without flattening them into generic text output.
• Multi-agent conversations. Bring several harnesses into one chat, or use side chats, ACP forks, and native subagent activity when available.
• Repeatable work. Create one-time or recurring schedules that run against a project and harness, with local run history.

A few things you can see

These are representative UI snapshots from the current artifact/E2E suite.

Readable activity history with grouped tool calls.	Interactive MCP Apps rendered directly in a turn.
Pop out a live result when the work needs more room.	Theme, language, density, and font settings are first-class.

Supported agents

Backchat currently ships registry entries for:

DeepSeek Harness can be connected in either form:

# Standalone ACP server
npm install -g @openma/deepseek-harness-acp
dsh-acp login

# Or install it into an existing dsh profile
dsh plugin --profile acp add -w @openma/deepseek-harness-acp

See openma-ai/deepseek-harness-acp for authentication, profile, model, permission, and MCP configuration.

Agent binaries are discovered or configured from Settings → Agents. The registry and setup layer are deliberately separate from the chat UI, so adding or updating a harness does not require changing the conversation model.

Quick start

Run from source

Requirements: Node.js 20 or newer and pnpm.

git clone https://github.com/openma-ai/backchat.git
cd backchat
pnpm install
pnpm dev

On first launch, open Settings → Agents to install or point Backchat at an ACP-compatible agent, then choose a project and start a chat.

Build and package

pnpm build          # production renderer/main bundles
pnpm package        # unsigned local installer for the current OS
pnpm package:dir    # unpacked app directory

Signed release packaging and notarization are documented in docs/releasing.md.

Development commands

pnpm typecheck
pnpm test:ci        # curated contract + release-regression lane (GitHub Actions)
pnpm test
pnpm test:e2e:fast
pnpm test:verify    # typecheck + full unit tests + fast E2E lane

Pull requests run test:ci and test:e2e:fast. Pushes to main also build an unsigned macOS DMG, then check that the packaged app can import its runtime and complete a first prompt. The website download button always points at https://github.com/openma-ai/backchat/releases/latest/download/Backchat-arm64.dmg ; tagged releases upload that stable filename next to the versioned DMG. GitHub's official Dependency Review, CodeQL, and Dependabot for Actions run alongside those gates.

The desktop shell is built with Electron, TypeScript, React, Vite, Tailwind CSS, and shadcn/ui. The ACP runtime is vendored from open-managed-agents and trimmed to the local desktop use case.

How the pieces fit together

1. Choose a project. A project is a working directory and the anchor for sessions and local artifacts.
2. Choose a harness. Backchat starts the selected ACP agent as a local child process and keeps its real configuration and capabilities visible.
3. Work in one surface. Conversation events, tool activity, MCP Apps, files, browser tabs, and terminals are projected into the same task.
4. Branch when useful. Continue in a side chat, fork an ACP session when supported, or follow native subagent activity emitted by the harness.

Status

Backchat is pre-release and under active development. The product shape, packaging, persistence format, and ACP integration surface may change before a stable release. It is ready for local evaluation and development, not yet a promise of backwards compatibility.

Issues and focused pull requests are welcome. For a useful bug report, include your OS, Backchat build, agent/harness, project type, and the smallest reliable reproduction.

Learn more

• Agent Client Protocol
• ACP session and subagent model
• Codex / ChatGPT plugin compatibility
• File-first storage RFC
• Release and signing guide