Monsters vs Goose
Goose and Monsters are open-source apps that run AI agents on your own computer with the model you choose. Goose is the larger and older project: one general-purpose agent with a desktop app, a CLI and an API, created by Block and now in the Agentic AI Foundation. Monsters is a young alpha built around a standing team of agents, a Markdown notes vault and safety controls.
Side by side
| Topic | Monsters | Goose |
|---|---|---|
| What it is | A desktop app with an orchestrator (Titan) and a squad of named agents, each with its own chat, terminal, permissions and model chain. | A general-purpose agent that runs on your machine, as a desktop app, a CLI and an API, for code, research, writing and automation. |
| License | Apache-2.0. | Apache-2.0. Created by Block, now in the Agentic AI Foundation. |
| Maturity | Alpha (0.1.0-alpha.20, 21 October 2026). Public since 20 October 2026. A young project with a small community. | Version 1.55.0 (16 October 2026), a release about every week. Repository since August 2024. About 55,500 GitHub stars. |
| Platforms | Linux (AppImage, .deb), macOS (Apple silicon and Intel, signed and notarized), Windows (installer, not yet signed with a public certificate). | Desktop app for macOS, Linux and Windows, plus a CLI. |
| Built with | TypeScript: a Node.js 24 core and an Electron app. | A Rust core and an Electron/TypeScript desktop app. |
| Models and providers | OpenRouter (free models to start), Anthropic, OpenAI, Google, Mistral and Groq with your own keys. A chain of models with fallback per agent. | 15+ providers, including Anthropic, OpenAI, Google, OpenRouter, Azure, Bedrock and Vertex AI. Any OpenAI-, Anthropic- or Ollama-compatible endpoint. |
| Local models | Yes: Ollama and LM Studio. | Yes: Ollama, LM Studio, Docker Model Runner. A model without tool calling can only chat. |
| Claude and ChatGPT subscriptions | Yes: Claude through Claude Code, ChatGPT and Codex through Sign in with ChatGPT. | Yes: Claude and Codex through ACP; ChatGPT and GitHub Copilot as native providers; Cursor through its CLI. |
| Many agents | Yes. The orchestrator delegates to the agents, in parallel. Whole projects and automations arrive as drafts you confirm. | Subagents that goose starts for a task, one after another or in parallel. Only in the autonomous mode, which is the default. |
| MCP | MCP client (stdio and HTTP) with a gallery of services. `monsters mcp` also serves your notes to Claude Desktop, Codex or Cursor. | 70+ extensions through MCP. |
| Notes and memory | An Obsidian-compatible Markdown vault under git: every agent edit is a commit you can undo. Keyword search (semantic when built from source). | Sessions kept in a local SQLite file, exportable as Markdown, JSON or YAML. A Memory extension. No notes vault. |
| Permissions by default | Terminal and file access are off until you grant them, agent by agent. | Four modes (autonomous, manual approval, smart approval, chat only). The default is Completely Autonomous. |
| Kill switch | Yes. One lever (the app, /stop on Telegram, `monsters kill`) aborts every run and kills the child processes. A program that detaches on purpose is not stopped. | Interrupt with Ctrl+C in the CLI, or with Send or a stop keyword in the app. No global switch documented. |
| Spending caps | Daily and monthly caps, per project too, for models with a known price. | No spending cap; a turn limit (`GOOSE_MAX_TURNS`) and a cost display in the CLI. |
| History and audit | Every action goes into a local action log. It is not tamper-proof against a process that can write your files. | Session history and application logs. |
| Remote use | Telegram bots (also one per agent, in groups), the terminal chat over SSH, and the desktop app as an SSH client of another computer. | A Telegram gateway (experimental) and an ACP server over HTTP and WebSocket on 127.0.0.1. |
| Scheduled work | Scheduled automations, proposed as drafts you confirm. | Cron schedules for recipes, from the Scheduler in the desktop app or the CLI. |
| Where data goes | Notes, logs and settings stay on your disk. Prompts go to the provider you choose (nowhere with Ollama or LM Studio). No cloud account, no usage analytics. | Runs locally. Prompts go to the provider you choose. Anonymous usage data is opt-in, asked on first use. |
Scroll the table sideways on a small screen.
Where Goose is ahead
- Maturity and community. Goose has been public since 2024, ships a release about every week and has about 55,500 GitHub stars. Monsters became public on 20 October 2026 and is a young project with a small community.
- Governance. Goose was created by Block and now belongs to the Agentic AI Foundation. Monsters is run by Saturday Company and its contributors.
- Reach. A full command line and an API you can embed, next to the desktop app. Monsters has a command line, but no API meant for embedding.
- Providers and extensions. More providers (Azure, Bedrock, Vertex AI and others) and a large catalog of MCP extensions.
- Recipes. Reusable tasks you can schedule and share.
Where Monsters differs
- A standing team. You get an orchestrator and named agents with their own chat, terminal, permissions and models, and they keep their identity between jobs. In Goose, subagents are temporary instances the main agent starts for a task.
- Notes you can read and undo. The memory is a Markdown vault you can open in Obsidian, and every agent edit is a git commit. In the pages we read Goose has no notes vault.
- Brakes from the start. Terminal and files are off until you grant them, per agent, and a small model (Instinct) brakes outgoing actions and asks you. Goose starts in its autonomous mode and lets you switch to approval modes.
- A kill switch and spending caps. One lever stops every run and child process, and daily and monthly caps apply per project, for models with a known price. Goose lets you interrupt a task and has a turn limit, but we found no global switch and no spending cap.
- Remote use as a main path: Telegram (also one bot per agent), the terminal over SSH, and the desktop app as an SSH client. Goose has an experimental Telegram gateway.
Pick Goose if
- You want the most established open-source agent app, with a large community and frequent releases.
- You want a CLI and an API to build on, not only a desktop app.
- You use a provider Monsters does not support yet, such as Bedrock or Azure.
- You like recipes and a wide catalog of MCP extensions.
Pick Monsters if
- You want a team of named agents with separate permissions rather than one agent that starts helpers.
- You want the agents’ memory as Markdown notes you can open in Obsidian and undo.
- You want a hard stop and spending caps built in.
- You want to drive agents from your phone with Telegram, or from another computer over SSH.
- You accept an alpha with a very small community.
Questions
Is Monsters an alternative to Goose?
For some uses, yes. Both are open-source apps that run AI agents on your computer with the model you choose. They follow different ideas: Goose is one general-purpose agent with subagents, available as app, CLI and API; Monsters is a standing team of agents with its own notes vault, spending caps and a kill switch. Goose is much more mature.
Can Monsters use the same models as Goose?
Mostly. Both work with Anthropic, OpenAI, Google, OpenRouter, Ollama and LM Studio, and both can use your Claude and ChatGPT or Codex subscriptions. Goose lists more providers, such as Azure and Bedrock.
Can I use Goose and Monsters together?
Yes, nothing stops you from installing both. They are separate apps with separate data, and both speak MCP, so the same MCP servers can be connected to either.
Sources
Read on 22 October 2026. Numbers such as stars and versions are the ones shown then. We found no mention of the features marked “not documented” in the pages listed here.