Exeora

Secure execution for AI agents, on any machine.

Connect Claude, ChatGPT, Cursor, VS Code or Claude Code to a real project on a server, a VM, a build box, a Raspberry Pi, or the laptop in front of you.

No port to open. No source code to upload. No tunnel to wire up.

cd the-project-you-want-to-serve
npx @exeora/cli connect

That one command signs you in, registers the machine, registers the directory, and prints an MCP URL. Point any client at it and leave connect running.

Hosted: exeora.dev · Docs: exeora.dev/docs · npm: @exeora/cli

Demo. Drop a terminal recording at docs/demo.gif when you have one (see docs/demo.gif.placeholder). Until then, the live product is at exeora.dev.

Why Exeora

AI agents are only as useful as the environment they can touch. Today you usually pick one of two bad options:

1. Expose the machine - open a port, run a tunnel, hope the URL stays private.
2. Upload the code - push the repo into a cloud sandbox that is never quite your machine.

Exeora is the third path. The CLI dials out to a gateway and holds a WebSocket open. Nothing ever dials in. Your files stay on the machine they already live on. The agent gets ten tools inside one project directory - not a shell on the whole box, and not a copy in someone else's cloud.

flowchart TD
    client["MCP client<br/><small>Claude · ChatGPT · Cursor</small>"]

    subgraph cloud["exeora.dev · Cloudflare"]
        gateway["Gateway<br/><small>OAuth 2.1 · MCP</small>"]
        relay["DeviceRelay"]
    end

    subgraph machine["your machine"]
        cli["Exeora CLI"]
        repo[("your repository")]
    end

    client -->|"Streamable HTTP"| gateway
    gateway --> relay
    cli -.->|"outbound WebSocket only"| relay
    relay -->|"tool.call"| cli
    cli --> repo

How it works

1. You run the CLI in the project directory. It opens an outbound connection to Exeora. No inbound port, no VPN, no ngrok config.
2. You authorize a client once. OAuth 2.1 with PKCE. Each project can be its own URL and its own token, so blast radius is a fact about the path, not a promise the model is asked to keep.
3. The agent works where the code already is - read, search, edit, run commands - under the policy you set. Revoke the machine from the dashboard and the socket closes immediately.

Compare

What you get

• No inbound network path - outbound HTTPS is the only requirement. Home routers, corporate proxies and cloud VMs with no public address all work unconfigured.
• Your code never leaves the machine - Exeora routes tool calls; it does not store the repository.
• A token per endpoint - a token minted for one project is refused at another. Ownership is checked in the database too.
• Paths stay in the project - every path is resolved before anything touches disk. .. and outward symlinks are rejected.
• Modes and command rules - read only, allow list, deny list, per-tool restrictions. Shell metacharacters are refused whenever a list is in force.
• Confirm before it runs - optional approval for edits and commands, in the conversation (MCP 2026-07-28) or on the terminal and dashboard.
• An audit log you can show someone - which tool ran, how it ended, how long it took. Never the arguments, never the output.
• Revoke and it stops - closing a machine kills the live socket that instant, not when a token expires.
• Bring your own client - Streamable HTTP, OAuth 2.1, PKCE, dynamic client registration. No plugin, no fork, no lock-in.

Tools

read_file · list_files · grep · edit_file · write_file · run_command · start_command · get_command_output · send_command_input · kill_command

On the account URL, three more for multi-project clients: list_projects · get_active_project · set_active_project.

Both URLs also carry get_agent_prompt, which reaches no machine: it hands back Exeora's own coding-agent instructions, for a client that arrived without any. See the agent prompt.

Full reference: exeora.dev/docs/tools.

Clients

Works with any MCP client that speaks Streamable HTTP and OAuth, including:

• Claude / Claude Code
• ChatGPT
• Cursor
• VS Code MCP
• MCP Inspector

Setup per client: exeora.dev/docs/clients.

The agent prompt

Claude Code and Cursor arrive knowing how to be a coding agent. claude.ai, ChatGPT and anything you wired up yourself do not, and it shows: they read files to find a symbol, overwrite a file they only half read, and treat a policy refusal as a problem to route around.

So Exeora ships a coding-agent prompt of its own, on four channels, because clients disagree about which they support:

• instructions in the MCP handshake. A short brief, arriving with nobody asking.
• The coding_agent prompt, for clients with an MCP prompt menu.
• The get_agent_prompt tool, for everything else, and for a model that decides to read first.
• exeora prompt, for whatever is not an MCP client at all.

exeora prompt > AGENTS.md     # or | pbcopy, or into a system prompt box
exeora prompt --account       # the variant for the account URL

It is about Exeora rather than about your codebase, so keep your own AGENTS.md exactly as it is. Read it at exeora.dev/docs/agent-prompt.

Install

Requires Node 22+.

# one-shot (recommended)
cd your-project
npx @exeora/cli connect

# or install globally
npm install -g @exeora/cli
exeora connect

Then add the printed URL to your client, for example:

claude mcp add --transport http exeora <the URL>

Documentation

Repository layout

Contributing

See CONTRIBUTING.md for local setup, tests and CLI releases.

Issues and pull requests are welcome. For vulnerabilities, email [email protected] - do not open a public issue.

License

AGPL-3.0-only.

If you modify Exeora and offer it as a network service, you must offer the corresponding source to the users of that service. Vendored third-party notices for the CLI live in packages/cli/LICENSE.