AkaraChen/cc-remote
Developer ToolsRun Claude Code as an ACP agent, keeping the model on a remote host while file and shell tools stay local.
What It Does
cc-remote turns Claude Code into an ACP agent whose model runs on a remote host while file edits, reads, searches and shell commands execute on your machine. It is built as a fork of claude-agent-acp with one key change: the claude process it starts lives on a server reached over SSH.
The split is deliberate. The server holds the model, system prompt and session state; your machine provides the built-in tools (Read, Write, Edit, Bash/PowerShell, Glob, Grep), your CLAUDE.md, git status, environment, and the MCP servers from your local Claude Code config and ACP client. The server only needs claude installed and logged in; your machine does not need it at all.
This suits developers who use ACP clients like Zed or paseo but want remote model compute, or who have a powerful server with Claude Code already authenticated. Any client that speaks ACP can use it.
Install and Connect
You need passwordless SSH access to the server (using a key, with ports and jump hosts in ~/.ssh/config) and Claude Code 2.1.248+ logged in there on a POSIX login shell. On your machine, Bun is required to build the binary.
git clone https://github.com/AkaraChen/cc-remote && cd cc-remote
bun install
bun run compile # builds a binary for this machine into dist/cc-remote
bun link # adds cc-remote to ~/.bun/bin
Tell cc-remote which host to use, then verify the connection:
mkdir -p ~/.config/cc-remote && echo devbox > ~/.config/cc-remote/host
cc-remote doctor
The host is resolved from --host, then $CC_REMOTE_HOST, then ~/.config/cc-remote/host. Anything ssh accepts works, and a custom command can be set via CC_REMOTE_SSH. If the server is not logged in yet, run cc-remote --cli auth login.
Add to an ACP Client
For Zed, add an agent in settings.json:
"agent_servers": {
"Claude (devbox)": { "command": "cc-remote", "args": ["--host", "devbox"] }
}
For paseo, you can add a provider under agents.providers in ~/.paseo/config.json, or install the bundled paseo plugin with paseo plugin install github:AkaraChen/cc-remote:paseo-plugin and configure the host through its settings screen. The plugin's "Check" button runs cc-remote doctor.
Runtime Behavior
Permissions are handled by your ACP client. Read, Glob and Grep never prompt; "Accept edits" covers Write, Edit and NotebookEdit inside the session directory. Bash approvals are granted per command prefix (e.g. npm test:*), not for all of Bash, and commands containing ;, &, |, redirects, $(…), backticks or newlines are never auto-approved. Approvals last until cc-remote exits.
Sessions are mirrored to ~/.claude/projects/ locally, so clients can list, replay, load and fork them. Resuming a session the server has not seen copies it over first. On the server, each local directory gets a working folder under ~/.cc-remote/ws/<hostname>/<local path>/. SSH connections share a ControlMaster by default with keepalive settings that fail within about a minute; Windows OpenSSH cannot share connections so each session opens its own.
Constraints
- Your project's
.claude/directory (commands, skills, agents, settings) is not sent to the server, though the server's user-level settings and skills still apply. - File checkpoints and rewind are recorded on the server and cannot see local edits made on your machine.
- Bridged MCP servers expose tools only, not resources or prompts.
- LSP, Monitor and worktree tools are disabled.
- On Windows the shell tool is PowerShell (pwsh if available, otherwise Windows PowerShell 5.1), and Grep requires
rgon PATH. - The repository declares no license.
HotRepo's Take
cc-remote appeared on HotRepo's new-project list on 2026-11-07, recording 22 stars and 3 forks gained in the roughly one-hour observation window from 04:45 to 05:45 UTC, reaching 137 stars total. The project arrived only days after its 2026-11-01 creation date, so the data shows a fast early spike rather than sustained adoption. No GitHub daily rank or HN discussion was recorded, so there is no broader trend signal to weigh. The technical design itself — splitting model execution from tool execution over a plain SSH stream using the SDK's tool channel — addresses a real constraint for ACP users with remote compute.
Written by AI from the project's README, updated . The project's own docs are the source of truth.