Quartet

A local-first AI agent workbench for chat, visual workflows, and scheduled automation.

English | 简体中文

Quartet brings the AI coding agents already installed on your machine into one browser-based workspace. It uses the Agent Client Protocol (ACP) as its agent boundary, so the same UI can run interactive conversations, compose repeatable multi-agent workflows, inspect live execution, and keep the resulting history on local storage.

Quartet is designed for a trusted shared instance running on a personal computer or another controlled environment. Multiple login accounts share the same workspaces and data while roles control available capabilities. It is under active development, and its APIs and stored data formats may change before a stable release.

Highlights

• One interface for multiple agents - maintains a catalog of built-in and custom ACP agents, probes their models and modes, and lets you switch agent, model, mode, and reasoning level from the conversation.
• Managed agent lifecycle - checks availability and versions and can run the catalog's explicit install, upgrade, and uninstall flows from Web or iOS.
• Interactive agent sessions - streams answers, thoughts, tool calls, images, and errors in real time; preserves session history and supports stopping, resuming, renaming, pinning, and sharing jobs. Follow-up messages sent while a job is running enter a durable queue shared by Web and iOS.
• Quiet completion hooks - a shared Shell hook can notify external systems after graph completion or an interactive turn. Interactive notifications are skipped by default while a visible Web or iOS page is already watching that Job live.
• Workspace-aware execution - organizes jobs around project directories, remembers per-workspace defaults, exposes files to the conversation, and shows the active Git branch.
• Visual graph workflows - connects Prompt, Clarify, Shell, If/Else, and Loop nodes on a canvas, with variables, concurrency controls, validation, hooks, live progress, and resumable human-in-the-loop steps.
• Scheduled automation - runs graph workflows from cron expressions with enable/disable controls, timeouts, concurrency limits, manual triggering, and links to the latest run.
• Local persistence and observability - stores workspaces, jobs, sessions, workflows, schedules, and usage statistics as local files; includes usage views by workspace, model, tool, and time range, plus daily token details and provider-reported cache hit rates.
• Messaging integrations - can receive and reply to tasks through Feishu/Lark and personal WeChat after they are configured in Settings.
• Native iOS client - Sophia, the SwiftUI app in ios/, talks to the same backend over your LAN for conversations, attachments, graph runs, scheduled tasks, and usage statistics.
• Agent-operated automation - ships quartet-cli plus workflow, schedule, and WeChat skills, so compatible coding agents can manage automation without modifying workflows authored by the user.

Screenshots

Home and workspaces

Agent conversation

Visual workflow editor

Usage statistics

Settings

iOS client (Sophia)

Recent tasks	New task

Agent conversation	Scheduled tasks

Usage statistics	Settings

Graph Workflows

Graph Workflows turn repeatable work into visual, executable processes instead of relying on one long prompt. A workflow is saved independently, can be bound to a workspace and working directory, and can be started manually or by a scheduled task.

Compose multi-agent processes

• Prompt nodes run an ACP agent with a selected model, mode, and reasoning level. Each node can start a new session or inherit context from an upstream agent session.
• Clarify nodes deliberately pause the workflow for human discussion. You can continue chatting in the node's session and resume the graph after the decision is ready.
• Shell nodes run project commands alongside agent work, making it possible to combine analysis with builds, tests, linters, scripts, or deployment steps.
• If/Else nodes route execution from variable-based conditions, while Loop nodes repeat a nested subgraph for a fixed count or until a condition is met.
• Graph branches can fan out and run concurrently, then join again. Workflow settings control concurrency, node and job timeouts, loop limits, and the maximum number of generated instances.

Variables carry structured results between nodes. Prompt and Shell nodes can publish named outputs for downstream prompts, conditions, and scripts. Before a workflow is saved or run, Quartet validates its topology, conditions, variable references, parallel writes, and unsafe concurrent session reuse, and reports all detected problems together.

Observe, control, and recover runs

• The canvas shows live node and edge states while each agent's conversation, reasoning, tool calls, command output, duration, and full error details remain available from the run.
• You can stop a run immediately, request a stop at the next step boundary, or cancel a pending step-stop.
• Failed, timed-out, interrupted, and step-stopped runs retain resumable state. Continuing a run preserves completed work and retries the unfinished portion, including the active round of nested loops.
• An active or resumable run can receive a new workflow version. Nodes that have not started use the new version, while running and completed nodes retain the configuration they actually executed.
• Per-node and end-of-run hooks can send notifications, record external status, or trigger follow-up scripts without changing the workflow result.

This makes Graph Workflows suitable for multi-agent code review, plan/implement/ verify pipelines, research with approval gates, repeated batch processing, and other tasks that need more control and visibility than a single chat session.

Code review demo

docs/demo/review-demo.json is a reusable multi-agent code review workflow. It alternates between two primary reviewers, asks the selected reviewer to double-check every finding, optionally sends the surviving issue list to another agent for adversarial verification, and then uses the inherited session to fix confirmed problems. Nested loops repeat the review and repair cycle so the result can converge over multiple passes.

The demo is a portable graph config and does not contain a machine-specific workspace ID or absolute path. Before running it, select your own workspace, adjust the Code, Doc, MultiWorker, AgentCheck, and Notice variables, and update the agent and model settings to match the ACP agents available on your machine. You can also validate and add it to the agent-managed workflow library with:

quartet-cli workflow validate --config-file docs/demo/review-demo.json
quartet-cli workflow create --name "Code Review" \
  --description "Iterative multi-agent review, verification, and repair" \
  --config-file docs/demo/review-demo.json

Scheduled Tasks

Scheduled Tasks run saved Graph Workflows automatically from standard five-field cron expressions. A schedule references a workflow rather than copying it, so every future trigger reads the latest saved workflow configuration.

• Choose the workflow, schedule name, cron expression, and optional workspace binding. An unbound schedule runs in the default workspace.
• Enable or disable a schedule without deleting it, and use Run now to test the same execution path immediately.
• Set a maximum number of concurrent runs to prevent unwanted overlap, plus an optional timeout for each scheduled execution.
• See the next trigger time, latest run time, latest status, run count, and the complete trigger error when a run could not be started.
• Open the latest generated Job directly to inspect the graph, agent sessions, tool activity, output, and errors just like a manually started run.

The scheduler runs inside the Quartet backend. Keep the backend running for scheduled automation; make web-watch can monitor the service and revive it when its port goes down.

Completion Hooks

Settings can define one default Shell hook for completion notifications and other side effects. Graph end nodes may opt into it, and every interactive turn considers it after completing, failing, or being stopped. By default, an interactive turn skips the hook while an authenticated Web or iOS page is visibly watching that Job. Hidden browser tabs, inactive or backgrounded iOS apps, public share readers, and internal subscribers do not count as viewers. A skipped notification is not queued for later; a later turn ending with no live viewer runs the hook normally.

The script receives QUARTET_* context variables including the trigger source, Job and session identifiers, run outcome, error text, latest assistant text, and viewer state. It runs from the Job workdir, while its temporary script file is created under $LOCAL_MEMORY/var/quartet/tmp/shell/ (or the operating system temp directory if LOCAL_MEMORY cannot be resolved). Hook output is diagnostic only, and hook failure never changes the task result.

iOS Client

ios/ contains Sophia, a native SwiftUI client for the same backend. It targets personal use over a LAN: no agent runs on the phone, and the app never reaches the host filesystem directly. Every capability goes through the same authenticated HTTP and SSE APIs the web UI uses, so both clients see one shared set of workspaces, jobs, and sessions.

• Connect and sign in - point the app at one Quartet address and log in with a Quartet username and password. Health and session checks run before the dashboard opens, a plaintext http:// endpoint requires explicit confirmation, the password is never stored on the device, and role permissions decide which actions appear.
• Recent tasks - a paginated job list with workspace filter, pin, rename, delete, and stop actions, plus a toggle that hides or reveals schedule-generated runs. Workspace and job summaries are cached, so the list still opens when the backend is unreachable and marks itself as out of date.
• Agent conversations - start a conversation by choosing workspace, agent, model, mode, and reasoning level, or continue an existing one. Answers, thoughts, tool calls, images, Markdown, tables, and individually copyable code blocks stream in over SSE, with token totals, elapsed time, and the active working directory and Git branch in the composer. Messages sent while an agent is busy remain in the server-backed queue across app restarts and stay synchronized with the Web client.
• Attachments - send photos, camera captures, and image files from the system picker; oversized images are compressed before upload.
• Presets and history - reuse message presets scoped to the current workspace, presets shared across all workspaces, and recently sent messages.
• Graph workflows - launch a saved workflow after reviewing and overriding its workspace, global run limits, initial variables, and per-node Prompt, Shell, Clarify, and If/Else settings. The run view shows progress, the execution trace, per-node agent sessions and shell output, and supports stop, stop after the current step, cancel a pending stop, resume, and completing a clarify discussion.
• Scheduled tasks - create, edit, enable, disable, delete, and manually trigger cron schedules bound to saved workflows, with next run time, run count, latest status, and the complete trigger error.
• Usage statistics - 7/30/90-day, all-time, and custom ranges with totals, duration, turns, daily token details, provider-reported cache hit rates, and rankings by workspace, model, and tool.
• Agent management - inspect built-in and custom agents, check availability and versions, run supported install/upgrade/uninstall flows, and configure environment variables, defaults, favorite models, and role-specific agents.
• Connection management - inspect the active endpoint and last successful sync, restart the web service, reconfigure the connection, or sign out and clear it.
• Background behavior - event streams stop whenever the app leaves the active foreground state and the server snapshot is re-read on return, so the UI does not silently display stale progress.

API errors keep the request method, URL, HTTP status, and full response body, and can be copied out of the app.

Building requires macOS with Xcode 26 or newer, an iOS 26 or newer target, and CocoaPods 1.15 or newer:

make pod-install   # install or refresh CocoaPods dependencies and the workspace
make build-ios     # build the device Debug configuration
make test-ios      # build the Simulator target without signing
make e2e-ios       # run the native XCUITest end-to-end suite in Simulator

Open ios/Quartet.xcworkspace rather than ios/Quartet.xcodeproj, select a signing team, and run. ios/README.md documents the current scope and verification boundaries in more detail.

Agent Support

Quartet includes the source for Eino, a standalone ACP agent that can be configured with Ark, OpenAI-compatible, Claude, DeepSeek, Gemini, Ollama, and Qwen model providers.

Quartet's built-in catalog currently supports the following ACP CLIs. The Agent management page can install, upgrade, uninstall, and validate entries whose catalog definitions provide those operations; custom ACP launch definitions can be added alongside them.

Agent	Required CLI	ACP command used by Quartet	ACP setup
Eino	eino-cli	eino-cli acp	Build and install from this repository with make install-eino-cli
TraeCLI	traex	traex acp serve	Provided by the CLI
Grok	grok	grok --no-auto-update agent stdio	Provided by the CLI
OpenClaw	openclaw	openclaw acp	Provided by the CLI
Claude Code	claude	claude-agent-acp	Requires the separate @agentclientprotocol/claude-agent-acp package
Antigravity	agy	antigravity-acp	Automatically installs the required antigravity-acp package and Bun
Cursor	cursor-agent	cursor-agent acp	Provided by the CLI
GitHub Copilot	copilot	copilot --acp --stdio	Provided by the CLI
Droid	droid	droid exec --output-format acp	Provided by the CLI
Kimi	kimi	kimi acp	Provided by the CLI
Codex	codex	codex-acp	Requires the separate @agentclientprotocol/codex-acp package
Kiro	kiro-cli	kiro-cli acp	Provided by the CLI
OpenCode	opencode	opencode acp	Provided by the CLI
KiloCode	kilocode	kilocode acp	Installable from the Agent catalog through npm
QCode	qoderclicn	qoderclicn --acp	Provided by the CLI

External tools, accounts, subscriptions, and authentication remain managed by their respective vendors.

You only need one working agent to start chatting. Unsupported or unavailable agents are skipped without preventing the rest of the application from loading.

Quick Start

Prerequisites

• Go 1.25 or newer
• Node.js >=22.18.0 <23
• npm >=10.9.0 <11
• Git, GNU Make, Bash, and common Unix tools such as lsof
• Credentials for at least one supported ACP agent; its CLI can be installed before startup or later from Settings > Agents

The current build and service scripts target Linux and other Unix-like environments.

1. Clone and configure local storage

git clone https://github.com/fanlv/quartet.git
cd quartet

export LOCAL_MEMORY="$HOME/.quartet-memory"
mkdir -p "$LOCAL_MEMORY"

LOCAL_MEMORY must be an absolute path. Quartet uses it as the root for its persistent configuration, jobs, sessions, workflows, schedules, uploads, and usage statistics.

2. Build and start Quartet

make web

On a normal source checkout, Quartet serves the built UI and API from a single backend process at:

http://127.0.0.1:8090

make web builds the frontend and backend, starts the backend as a detached process, and writes logs to /tmp/quartet-backend.log.

3. Open the UI

On first launch, create the administrator account; later visits require a Quartet login. Create or select a workspace, choose an available agent and model, then send a message. Agent installation, environment variables, favorite models, defaults, and role-specific agents are managed under Settings > Agents.

External ACP Agents

The easiest setup path is Settings > Agents > Install & Upgrade, which shows the exact catalog-defined commands before running them and retains complete results. You can also install and authenticate an external agent manually; in that case, ensure its CLI and ACP adapter are available on the same PATH used to start Quartet, then restart the backend.

Claude Code and Codex do not expose the ACP commands used by Quartet through their main CLI packages alone. After installing the claude or codex CLI, install the corresponding ACP adapter separately:

npm install -g @agentclientprotocol/claude-agent-acp
npm install -g @agentclientprotocol/codex-acp

Quartet requires both claude and claude-agent-acp for Claude Code, and both codex and codex-acp for Codex, to resolve from the backend's PATH.

Service Commands

Command	Description
make build	Build CLI, Eino, backend, and frontend artifacts
make web	Build the UI and backend, then start or restart the detached web service
make web-status	Show backend and watchdog status
make web-logs	Follow /tmp/quartet-backend.log
make web-stop	Stop the watchdog, backend, and orphaned quartet-web processes
make backend-stop	Stop only the backend and leave the watchdog running
make web-watch	Start a detached watchdog that revives the backend if its port goes down
make web-watch-stop	Stop the watchdog without stopping the backend
make web-watch-logs	Follow /tmp/quartet-watchdog.log
make build-frontend	Rebuild the SPA into static/ without restarting the backend
make build-web	Build bin/quartet-web without restarting the service
make build-cli	Build bin/quartet-cli
make build-eino-cli	Build bin/eino-cli
make install-eino-cli	Build and install the standalone Eino ACP agent
make install-project-tools	Install quartet-cli and all project skills
make test-go	Run all Go tests
make test-web	Run frontend component tests
make lint-web	Run frontend ESLint checks
make e2e	Run frontend Playwright tests
make build-ios	Build the iOS app on macOS with Xcode
make test-ios	Build the iOS app for Simulator without signing
make pod-install	Install or refresh the iOS CocoaPods dependencies
make e2e-ios	Run the native iOS UI tests in Simulator

Configuration

Variable	Required	Description
LOCAL_MEMORY	Yes	Absolute path used for Quartet's persistent data and runtime state
QUARTET_LISTEN_ADDR	No	Overrides the default listen address
QUARTET_CORS_ORIGINS	No	Comma-separated cross-origin allowlist; unset means same-origin only
QUARTET_TRUSTED_PROXIES	No	Comma-separated reverse-proxy IPs/CIDRs trusted to supply client-IP headers; defaults to loopback only; use none to disable
QUARTET_LOG_LEVEL	No	Initial log level: debug, info, warn, or error
QUARTET_LOG_HTTP_BODY	No	Log HTTP request and response bodies; disabled by default and may expose secrets
QUARTET_STATIC_DIR	No	Built frontend directory; defaults to static
QUARTET_CERTS_DIR	No	Directory containing cert.pem and key.pem; defaults to certs
QUARTET_MAX_ACP_AGENTS	No	Maximum number of cached ACP agent instances
QUARTET_ACP_PROBE_CONCURRENCY	No	Maximum concurrent ACP capability probes
QUARTET_MESSAGES_CACHE_BYTES	No	In-memory message-history cache budget; 0 disables it
QUARTET_BASE_URL	No	Backend URL used by quartet-cli; defaults to http://127.0.0.1:8090

Without certificates, Quartet binds to 0.0.0.0:8090 over HTTP. When both cert.pem and key.pem exist in the certificate directory, it enables HTTPS and defaults to 0.0.0.0:443; it also exposes a loopback-only plaintext listener at 127.0.0.1:8090 for quartet-cli and local workflow scripts. QUARTET_LISTEN_ADDR overrides the backend address but does not change whether certificate-based TLS is enabled.

Quartet always protects private APIs with user login sessions. On first start, use the Web setup page to create the first administrator. Administrators can then create users and assign roles. Web, iOS, and quartet-cli all authenticate with cookies; there is no shared API token. See the permissions and access-control documentation for the complete boundary.

Data and Privacy

Quartet's own state is file-based and local:

$LOCAL_MEMORY/
├── quartet/
│   ├── config/       # auth, prompts, workflows, schedules, and message presets
│   ├── data/         # settings, Agent catalog, workspaces, jobs, uploads, IM, and shares
│   └── usage-stats/  # month-sharded usage statistics
└── var/quartet/
    ├── state/        # sessions and schedule state
    ├── cache/        # reconstructable caches
    └── tmp/          # process-owned temporary files; hooks use tmp/shell/

Writes to important records are atomic. Backing up LOCAL_MEMORY backs up the complete state of the Quartet instance. Login accounts share its workspaces and business data; roles control capabilities rather than providing data isolation.

Messages and files are still sent to whichever agent and model provider you select. Review that provider's privacy policy and the permissions granted to its local CLI.

Architecture

flowchart LR
    UI["React + Vite web app"]
    IOS["SwiftUI iOS app"]
    CLI["quartet-cli + project skills"]
    API["Go + Hertz backend"]
    ACP["ACP agent processes"]
    DATA["Local file storage"]
    IM["Feishu / Lark and WeChat"]

    UI <-->|HTTP + SSE| API
    IOS <-->|HTTP + SSE| API
    CLI <-->|authenticated HTTP| API
    API <-->|ACP over stdio| ACP
    API <-->|atomic reads and writes| DATA
    IM <-->|messages and media| API

Loading

HTTP handlers validate and authorize requests, business services own behavior, repositories own persistence, and types/path owns the storage layout. All agents, including Eino, enter through the same ACP session and event pipeline.

Path	Purpose
cmd/web	Web server, API routes, middleware, and application assembly
cmd/quartet-cli	Authenticated workflow, schedule, workspace, job, agent, and WeChat CLI
cmd/eino-cli, einocli	Bundled standalone Eino ACP agent
types/model	Shared request, response, and domain models
types/path	Canonical paths for configuration, business data, and runtime state
repository	Local persistence
services	Auth, agent, job, graph, schedule, workspace, IM, skills, and statistics behavior
pkg	ACP, messaging, file storage, logging, and common infrastructure
web	React frontend
ios	Native SwiftUI client (Sophia) for personal LAN use
skill	Workflow, schedule, and WeChat skills driven by quartet-cli

CLI and Project Skills

Install quartet-cli and all three project skills for supported coding agents:

make install-project-tools

The command installs quartet-cli to ~/.local/bin by default and registers:

• quartet-workflow for creating, inspecting, validating, updating, deleting, and running workflows in the agent-managed library. User-authored workflows remain read-only to the CLI.
• quartet-schedule for creating and operating cron schedules backed by saved Graph Workflows.
• quartet-wechat for listing connected iLink accounts and sending proactive text messages through the durable WeChat outbox.

The CLI also lists workspaces and installed agents and can inspect or stop Jobs. To install only one project skill, set its name explicitly:

make install-skill SKILL_NAME=quartet-workflow

quartet-schedule and quartet-wechat are the other valid project skill names.

Development

make build           # Build all application artifacts
make build-all       # Build all Go applications
make build-cli       # Build quartet-cli
make build-eino-cli  # Build the standalone Eino ACP agent
make install-eino-cli # Build and install Eino to INSTALL_BIN_DIR
make build-frontend  # Type-check and build the React application
make test-go          # Run all Go tests
make test-ios        # Build the iOS Simulator target on macOS without signing
make e2e-ios         # Run native iOS UI tests in Simulator
make test-web        # Run frontend component tests
make lint-web        # Run frontend ESLint checks
make e2e             # Run Playwright end-to-end tests
make test            # Run Go tests, frontend tests, and E2E tests
go test ./...        # Run Go tests

Frontend-only commands are available from web/:

npm run dev
npm run build
npm run lint
npm test

Contributing

Issues and pull requests are welcome. Before submitting a change, keep behavior within the existing service boundaries and run the checks relevant to the code you changed.

License

Quartet is licensed under the Apache License 2.0.