Limitations and roadmap

Agentbits is an MVP, version 0.1.0. It already runs real work with real guardrails, but some things are missing and some work differently than you might expect. This page lists them plainly, so you can decide what to rely on.

What Agentbits 0.1 does not do

Installation and platforms

• No published installer yet. You can run Agentbits from source or build the Windows installer yourself. See Install.
• No signed installers. When installers are published, Windows SmartScreen and macOS will warn you on first launch.
• No tested macOS package and no Linux package. Windows is the main target. Running from source works on Windows, macOS and Linux.
• No self-update. Install a newer version over the old one. Your data is kept.
• Desktop only. There is no web or mobile version.

AI providers

• Five providers only: Claude Code CLI, Codex CLI, Anthropic API, OpenAI API and Demo. No Gemini and no local models yet.
• Agentbits doesn’t bundle, install or sign in to the CLIs. You install Claude Code or Codex CLI and sign in yourself in a terminal.
• No signing in with a Claude subscription inside the app. For Claude, use an API key or your own Claude Code sign-in.
• API agents can’t edit files or run commands. Anthropic API and OpenAI API agents work with text and Agentbits’ own tools (connections, files, skills, apps). Changing code in a project folder needs a CLI provider.
• Codex CLI agents have no connection or file tools. Files reach Codex only through the prompt and as images.
• No spending limits. Agentbits counts tokens, but you can’t cap tokens, time or money per task or project. A cost in dollars appears only when Claude Code reports it. See Providers.

Agents and automation

• Agents can’t use the built-in browser. The Browser permission exists but is not applied yet. Web access works only through a CLI provider with the Network permission.
• No scheduled or recurring tasks, and no deadlines. Tasks come from the chat, from you, or from instructions. The autopilot only takes the next task in line.
• No conflict handling when several agents edit the same folder at once. Use git branches.
• No git integration. Agentbits doesn’t create branches, commits or pull requests. An agent with the Terminal permission can run git itself.
• No operating system notifications. You get in-app toasts and badges only.

Files, memory and reports

• No OCR. Scanned PDFs without a text layer are only partly read, and the app says so.
• Legacy .doc, .xls, .ppt and .rtf files are stored but not read.
• Agents never change or delete your attachments. They only read them and can save new text files.
• Memory is full-text search, not vector (embedding) search.
• No custom reports. Reports offer statistics, agent performance, task reports, the daily report and token usage.

Apps and skills

• Apps (MCP): tools only. MCP resources, prompts and sampling are not supported.
• Agentbits doesn’t install MCP servers or the Node.js or uv they may need. You give the server’s command or address.
• Skills come from public GitHub repositories, folders, .zip or .md files. Private repositories are not supported.

Data and accounts

• One computer, one user. No cloud sync, team accounts or remote access. Everything stays in a local database.
• Backups cover the database only, not your project folders.

Known rough edges

• Pause doesn’t freeze CLI agents. For Claude Code and Codex CLI, pause holds the output stream. The process may keep working in the background.
• Terminal permission is not sandboxed by the operating system. An agent with Terminal and Network can do what your user account can do, including using credentials stored on your machine. See Security.
• More than 10 workers in one project share desks in the office.
• Stored text keeps its language. Changing the app language doesn’t translate chat messages, status notes, audit records or task reports already saved.
• Less-tested drivers. MySQL and SQL Server connections have had less testing than PostgreSQL and SSH.
• App sign-in (OAuth) for MCP servers has been tested only with a demo authorization server, not with real services.

Roadmap

These items are planned. There are no dates.

• Published, checked installers on GitHub Releases (Windows first, then macOS)
• More providers, for example Gemini and local models
• Agents using the browser as a tool (the Browser permission)
• Connection and file tools for Codex CLI agents
• MCP resources, prompts and sampling
• Skills from private GitHub repositories
• Spending limits per task or project
• Conflict handling for agents editing the same folder
• Vector (embedding) memory and custom reports
• OCR and legacy Office formats
• More interface languages (community translations)

Want something that isn’t here? Open an issue on GitHub and describe the problem you want to solve.