一个工作空间
协同多个 AI Agent

支持连接兼容 ACP 协议的 Agent，
包括 Claude Code, Codex, Grok, Copilot 等。

Copied!
curl -fsSL https://caelis.dev/install.sh | sh

DESIGN PHILOSOPHY

Leverage agent specialization. Maintain a unified contextual session.

Caelis treats ACP as an open bridge between agent runtimes, fostering seamless context handoffs in your terminal rather than locking intelligence into isolated chat UIs.

01

Local-first Runtime

All agents execute natively on your machine. Your codebase, commands, and runtime boundaries stay entirely in your control.

02

Open Agent Context Protocol

Connect and orchestrate different ACP-compatible agents in one session. Expand your workspace without closed platform silos.

03

Secure Execution Boundaries

Run agents safely inside custom sandboxes. Strict permissions and manual approval workflows prevent malicious shell commands.

04

Cross-Platform Support

Consistently run Caelis runtime and sessions across macOS, Linux, and Windows environments.

05

Ecosystem Compatibility

Reuse your existing skills, plugins, MCP servers, and ACP clients. Migrate tools with zero vendor lock-in.

06

Open Source & Extensible

Licensed under Apache-2.0. Standardize agent orchestration with an active, extensible community runtime.

A shared agent workspace

caelis - terminal session
>_ CAELIS (v0.21.0)
model:claude-3-5-sonnet [local] workspace:~/project tip:type / for command list
Just run a simple command for demonstration.
> The user wants me to run a simple command for demonstration. This is a straightforward request - no skills apply here. Let me just run a simple command.
• Ran echo "Hello from Caelis! 🚀 Current date: $(date)"
└ Hello from Caelis! 🚀 Current date: Sun Jul 5 12:20:37 CST 2026
> The command ran successfully. Let me share the result with the user.
• Command executed successfully! Here's the output:
Hello from Caelis! 🚀 Current date: Sun Jul 5 12:20:37 CST 2026
The environment is working — shell commands run, return exit codes, and produce output. Let me know if you'd like to do something more interesting!
3.5s
@amir Briefly introduce yourself.
> The user is sharing a "Caelis" shared dialogue context. This appears to be some kind of mu...elf briefly. I don't need to use any tools for this - it's just a conversational response.
• Hi! I'm Claude, an AI assistant by Anthropic. I'm here to help with software engineering tasks – debugging, writing code, exploring codebases, planning features, and more.
I can see the shared context showing a successful command demo in your workspace at ~/project. What would you like to work on?
3.5s
~/project claude-3-5-sonnet [local]
> Type your message, @agent, #path/to/file, or $skill
auto-review shift+tab
7.2k / 1.0m • 0%

FAQ

Common questions

What is Caelis?

Caelis is an open-source, local-first runtime for AI agents. It manages local workspaces, session context, tool execution, and permission boundaries, and can connect compatible agents and clients via ACP.

Can Caelis run independently?

Yes. Caelis includes a built-in local Agent Runtime and does not require third-party clients to execute command tasks via TUI or Headless CLI.

Is ACP mandatory?

No. Caelis runs local agents independently out of the box. ACP is primarily used to connect compatible external agents, editors, or clients to share a unified workspace and context.

Why not just use Claude Code directly?

Claude Code, Codex, and OpenCode are excellent agents. Caelis focuses on the runtime layer: managing workspaces, context sharing, sandbox isolation, permission prompts, and multi-agent coordination. Instead of replacing any specific agent, Caelis enables them to collaborate in a single secure environment.

Does it support existing Plugins, Skills, and MCP ecosystems?

Yes. Caelis reuses community capabilities rather than inventing closed ecosystems. You can connect MCP servers, call Skills, and utilize existing plugin marketplaces or tool extensions to migrate your workflows with low friction.

Is my data kept locally?

Yes. Caelis is local-first. Your local workspaces, session transcripts, execution histories, and runtime configurations are stored on your machine. External model APIs are called only during inference depending on your model choices.

How does Caelis ensure secure agent execution?

Caelis runs agents in sandboxed environments by default and uses Guardian for automated tool review. Low-risk operations execute automatically, while actions requiring host access or exceeding sandbox boundaries trigger privilege prompts. Developers can also toggle optional manual approval workflows.

Does it support Windows, macOS, and Linux?

Yes. Caelis supports macOS, Linux, and Windows. Depending on your platform, you can install Caelis via shell script, PowerShell, or global npm package.

Is Caelis open source?

Yes. Caelis is open source and licensed under the Apache-2.0 License. Code is hosted on GitHub, and we welcome issue reports, pull requests, and ecosystem contributions.

Is a native GUI planned?

Yes. Caelis currently supports TUI, Headless CLI, and ACP Runtime. Native GUI is on our active roadmap and will connect as a new interface to the same local runtime, rather than being a separate product.