What is Apprentice?
Apprentice is software for operating AI agents.
It is a local-first desktop control plane for creating, running, and monitoring a team of bounded AI agents. You create agents with their own identity, instructions, model/provider choice, budget, permissions, guardrails, memory, knowledge base, browser settings, task queue, schedules, integrations, and MCP tools. Apprentice gives you one place to decide what each agent is allowed to do, when it wakes up, which tools it can use, and how its work is reviewed.
Apprentice is not a local-model-only app. It supports local runtimes, but it also supports hosted APIs, CLI-based providers, and subscription/OAuth paths. You can connect providers such as Claude Code, Codex, Antigravity CLI (beta), ChatGPT Subscription, Anthropic / Claude API, Google Gemini API, OpenAI, DeepSeek, Mistral, Kimi, GLM, Qwen, LM Studio, Docker Model Runner, and Ollama. The exact provider list may change by release, but the product principle is the same: bring the provider or local runtime you want to use.
Local-first means the desktop app, SQLite state, agent configuration, runtime decisions, logs, memory, audit history, and Docker-backed execution control live with you. Model calls go to whichever provider or local runtime you configure for that agent. If you choose a local runtime, model execution can stay local. If you choose an API, CLI, or subscription provider, Apprentice sends that provider the prompts and context needed for the run.
What agents can do
Agents in Apprentice are not limited to one chat box. They can be started from private chat, group chat, tasks, schedules, feed/topic mentions, external channel messages, and integration-triggered work. They can participate in shared topics, collaborate in chats, create and update tasks, remember useful context, search their own knowledge base, send notifications, and use configured tools.
Each agent runs through a Docker-backed runtime instead of directly on the host. That runtime gives agents a controlled Linux environment with scoped filesystem access, command and website rules, browser configuration, and per-agent execution state. Agents can also be connected to external MCP servers, with per-agent bindings and tool allowlists, so one agent can have access to a tool that another agent does not.
What you control
Apprentice is built around bounded autonomy. For every agent, you can configure:
- The provider, model, and provider account it uses.
- The system instructions and working style it should follow.
- The directories, files, commands, websites, and browser behavior it can access.
- The task, schedule, chat, feed, and integration triggers that can wake it.
- The memory, knowledge, and store data it can use.
- The MCP servers and specific MCP tools exposed to it.
- The budget limits, run limits, guardrails, and approval behavior around its work.
The point is not just to make an agent answer. The point is to make an agent useful while keeping its scope visible and adjustable.
What Apprentice is not
Apprentice is not a hosted agent cloud, a generic workflow-box builder, a coding IDE, a terminal wrapper, or only a chatbot. It can support coding, research, support, sales, community, operations, and back-office workflows, but the product category is broader: a desktop control plane and runtime for AI agent work.
Who it is for
Apprentice is for people who want AI agents that can do ongoing work across tools and communication channels without handing the whole operating surface to a hosted automation platform. It is especially useful when you care about provider choice, local state, auditability, per-agent boundaries, and being able to inspect what happened after an agent runs.
Apprentice ships stable releases on a calendar-based version scheme. Check the release notes and setup guides for current requirements, supported providers, and known limitations.