MindFS

English | 简体中文 | Official Site | Discord | Twitter | 【wechat group】 | Competitive Analysis | Competitive Analysis PPT

AI Agent Remote Access Gateway · Result Visualization

Access your personal AI agents and workstation data anywhere, anytime through MindFS.

---

Screenshots

---

Features

Agent Sessions

• Multi-Agent support: Claude Code · OpenAI Codex · Gemini CLI · Grok · Cursor · GitHub Copilot · CodeBuddy · Cline · Augment · Kimi · Kiro · Qwen · Qoder · OMP · Pi · Hermes · DeepSeek Harness (DSH) · Reasonix · OpenCode · OpenClaw — installed agents are detected automatically.
• Real-time streaming: Token-by-token output pushed to the browser; tool calls, thought traces, permission prompts, and remaining context-window capacity are rendered live as structured, collapsible cards.
• Flexible switching: Switch agents or models mid-session; all agents share the same context — no need to re-explain the background.
• Session search: Search by session title or conversation content, then jump straight to the matched session and snippet.
• External session both-way import&synchronization: Browse existing sessions from supported agent CLIs, import one into MindFS, and continue it as a native MindFS session. MindFS session can resumed in agent cli too. Bidirectional synchronization is also possible.
• Binding persistence and recovery: MindFS persists the mapping between its internal session and the underlying agent session, so the link can be restored after service restarts and follow-up messages continue on the same agent session when available.
• Rich media input: Attach files and images directly in your messages.
• Multi-device sync: Access the same instance from multiple devices simultaneously with live session sync.
• Configuration Management & Switching: Agents configurations / API providers can be backup / added and switched with one click, making it easier to move between multiple accounts or API keys.
• Subagents: Codex/claudecode subagents are automatically discovered and displayed.
• session fork: Session can forked from historical replies.
• Scheduled tasks: Trigger agents to run tasks at specified times.
• Codex Remote Login: Login to Codex remotely via the /login command.
• Built-in Token Stagion: Real-time balance display and one-click activation.
• Token Usage data: Record and display token usage and context-window usage.

Task Board

• Task Orchestration: Automatically break down complex goals into subtasks with dependencies, coordinate parallel execution across multiple agents, and complete the workflow with an overall acceptance review in the parent conversation.
• Concurrent Execution: Run multiple tasks in parallel, with each task isolated via its own worktree.
• Task Templates: Customize task stages within templates. Each stage supports configuration of the agent, model, planning mode toggle, preset prompts, and more.
• Deep Linking: Dynamic, interactive associations between tasks, worktrees, sessions, and files.
• Quick Creation: Instantly spin up tasks from templates, with support for file uploads, images, and skill inputs.

File Access

• Multiple projects: Manage several directories at once; sessions are organized per project and stay independent.
• Self-hosted data: Conversation history, file metadata, and view config are stored in the project's .mindfs/ by default. The sidebar menu can place metadata for new projects under ~/.mindfs/<rootId>/; an existing project-local .mindfs/ is always reused.
• File tree browser: Full directory navigation with file preview; Markdown, images, and code all have dedicated renderers.

Interaction

• / slash commands: Type / to trigger a command palette and quickly run preset operations.
• @ file references: Type @ to trigger path completion and attach any file as context for the agent.
• # quick prompts: Type # to trigger your saved prompt shortcuts inline.
• Bidirectional file–session linking: Jump from a file to the session that created it, or from a session to all files it touched.
• Android, Browser app (PWA): Install to desktop or mobile home screen for a native-like experience — no app store required.
• Mobile-optimized UI: Bottom action bar within thumb reach, independent panel swipe navigation, input box adapts to the soft keyboard.
• Notification Push: Notifications are delivered via Web Push upon session status changes (iOS requires adding the page to the home screen).
• User Habit Adaptation: Support swapping the left and right sidebars, and switching between single-project and multi-project session lists.

Access Modes

• Local mode: Accessible in the browser on the local machine immediately after startup — no account or configuration needed.
• Relay remote mode: Access your local instance from anywhere on the public internet without opening firewall ports, via an encrypted tunnel through a9gent.com. Click the bind button in the local UI to activate.
• Private channel: Use a private network (e.g. Tailscale) and access directly via ip:port.
• End-to-end encryption: Sessions and files can be protected with end-to-end encryption.

Plugin System

• Custom views: A plugin is a custom view for a file, following the pattern: receive file content → parse → render UI.
• Agent-generated plugins: Tell the agent "implement a txt novel reader" and it generates the plugin — all txt files are then displayed as a reading experience.
• Interaction loop: Plugins can register action buttons that send structured commands to the agent, completing the loop: customize plugin → browse file → agent interaction.

Command Execution

• Card output: Command results are displayed as cards for clearer reading.
• History suggestions: Matching previous commands automatically appear as suggestions for quick input.
• Screen-width adaptation: Command output adapts to the screen width for a friendlier result view.
• Selectable shell type: Choose the shell used for command execution, so Windows users do not need to worry about shell type mismatches.
• Session persistence: Each session gets a long-lived shell, making tmux-like command continuity easier.

Remote Access to Local Services

• One-Click Exposure: In Relay mode, configure the address of your local service in MindFS to expose it to the public network with one click.
• Public Domain Name: Access local services directly via a unique public domain name.

Installation

• Single binary: The production build is a statically compiled binary with all web assets embedded, and the install package is under 10 MB.
• Zero dependencies: No Node.js, Docker, or daemon manager required on the host.
• Cross-platform: macOS (Intel + Apple Silicon), Linux (x86-64, ARM64, ARMv7), Windows (x86-64, ARM64).

---

Quick Start

Prerequisites

MindFS does not include any AI model — you need at least one Agent CLI installed locally. Choose what works for you:

MindFS also collects commonly used agents in the local UI. Open the file-tree menu, choose Agent Install & Update, then use the generated install/update commands for your platform. The commands are launched in MindFS command mode so you can review and run them from the same workspace.

For DSH, the generated commands install the OpenMA ACP adapter into a dedicated mindfs-acp profile. Configure the provider, API key, and default model in dsh web; MindFS uses the same per-user $DSH_HOME/settings.yaml and $DSH_HOME/.credentials.yaml. The dedicated profile isolates the ACP composition only—it does not replace or hard-code the user's DSH Web configuration.

Once an agent is installed, start MindFS and interact with it through the browser.

Install

macOS / Linux

curl -fsSL https://raw.githubusercontent.com/a9gent/mindfs/main/scripts/install.sh | bash

Custom install path:

curl -fsSL https://raw.githubusercontent.com/a9gent/mindfs/main/scripts/install.sh | bash -s -- --prefix your/path

Windows (PowerShell)

irm https://raw.githubusercontent.com/a9gent/mindfs/main/scripts/install.ps1 | iex

The install script auto-detects your OS and architecture, reads the latest version from the first line of release-notes.md, then downloads the matching binary from GitHub Releases. release-notes.md keeps release history with the newest entry at the top; make release TAG=v1.2.3 commits and pushes it when changed, then uses only the top entry as the GitHub release notes.

Uninstall

The uninstall command removes the installed binary, bundled web assets, bundled default agents.json, and the PATH entry added by the installer. User config and project .mindfs/ data are kept by default.

macOS / Linux

installer="${TMPDIR:-/tmp}/mindfs-install.sh"
curl -fsSL https://raw.githubusercontent.com/a9gent/mindfs/main/scripts/install.sh -o "$installer"
bash "$installer" --uninstall

Windows (PowerShell)

$Installer = Join-Path $env:TEMP "mindfs-install.ps1"
Invoke-WebRequest -Uri "https://raw.githubusercontent.com/a9gent/mindfs/main/scripts/install.ps1" -OutFile $Installer
& $Installer -Uninstall

To also remove user-level MindFS config and logs, add --purge on macOS/Linux or -Purge on Windows. Project .mindfs/ directories are never removed automatically.

Build from source (requires Go 1.22+, Node.js 20+)

git clone https://github.com/a9gent/mindfs.git
cd mindfs
make build      # output: ./mindfs

Run

mindfs                        # manage current directory
mindfs /path/to/your/project  # manage a specific directory
mindfs -addr :9000 /path/to/your/project  # custom port

Open http://localhost:7331 in your browser.

HTTPS (TLS)

Enable HTTPS with a self-signed certificate (auto-generated and reused across restarts):

mindfs -tls
mindfs -tls -addr :9000 /path/to/your/project

Open https://localhost:7331 in your browser. The auto-generated certificate includes SANs for localhost, 127.0.0.1, ::1, and all non-loopback interface IPs, so LAN clients can connect without certificate name mismatches. Certificates are stored under the user config directory (e.g., ~/.config/mindfs/ on Linux).

Use custom certificate and key files:

mindfs -tls -cert /path/to/cert.pem -key /path/to/key.pem

MindFS automatically detects the availability of installed agents. This usually takes about one minute.

Enable Remote Access (Optional)

1. Open MindFS in local mode and click the bind button in the bottom-left corner.
2. Log in to a9gent.com and confirm the binding.
3. Open your node — it is now accessible from any device.

Custom ACP Agents

MindFS can load an extra agents.json for agent CLIs that speak the ACP protocol. This is useful when you are testing a new agent or want to keep project-specific agent definitions outside the bundled defaults.

{
  "agents": [
    {
      "name": "my-agent",
      "brief": "Short description shown in the install/update list.",
      "command": "my-agent",
      "protocol": "acp",
      "args": ["--acp"]
    }
  ]
}

Start MindFS with the extra config:

mindfs -agent-config /path/to/agents.json

The extra file can add new agents or override a bundled definition with the same name.

CLI Reference

mindfs [flags] [root]

root is the directory to manage. If omitted, MindFS opens without adding a directory.

By default, mindfs starts or reuses a background service and opens the browser without registering automatic startup. Run mindfs -autostart once to enable it for the current user. Once enabled, normal terminal launches refresh a private snapshot of the visible environment and the existing startup entry. At automatic startup, MindFS restores that snapshot and then loads the recorded shell's rc/profile so newer shell values take precedence. Pass -autostart=false to remove the startup entry and environment snapshot.

MindFS uses a launch agent on macOS, a systemd user service on Linux, and the current user's startup registry entry on Windows. The environment snapshot can contain credentials and is stored under the platform user config directory (os.UserConfigDir()/mindfs) with user-only permissions.

Common Commands

mindfs
mindfs /path/to/project
mindfs -addr :9000 /path/to/project
mindfs -foreground /path/to/project
mindfs -status
mindfs -version
mindfs -update
mindfs -uninstall
mindfs -stop
mindfs -restart
mindfs -remove /path/to/project
mindfs -agent-config /path/to/agents.json

Flags

---

Contributing

Pull requests are welcome. For larger changes, please open an issue first to discuss the approach.

wechat-group

---

License

AGPL v3

Friend Links

Linux.do