Skip to main content

code-server

VS Code in the browser, over the workspace, from any device.

Access​

  • Browser: https://<machine>.<tailnet>.ts.net over Tailscale (Core), or https://code-server.EXIST_DOMAIN with the Caddy add-on
  • Password: CODE_SERVER_PASSWORD in services/code-server/.env

Enable​

Part of Core, so it ships on. If you turned it off:

EXIST_IS_SERVICES_CODE_SERVER=true

Then ./existential.sh && docker compose up -d from the repo root.

  • First boot installs code-server, five extensions, Claude Code and opencode: a couple of minutes, unhealthy until done.
  • About 210 MB idle; 1.3 GB peak during the first install. The 2 GB limit is headroom.

Shared workspace​

It mounts workspace/, the same tree the automation daemon's agents see, so you and an agent can edit the same files at once.

Seeing your flows​

Open any graph in .decree/graph/ and run Markdown: Open Preview to see it as a diagram (Mermaid extension, installed on boot).

AI in the terminal​

The terminal has claude, opencode and pi. Talking to one is how you make new automations. Machines have their own Claude login: Decree.

claude is Claude Code. Log in once; to keep the conversation going from your phone, start a remote-control session in tmux:

tmx -d claude claude remote-control   # starts it, unless one is already running
tmx claude                            # attach to it

tmx never starts a second copy.

opencode is pointed at the stack's chat model on every boot (EXIST_MODEL_URL_CHAT, else EXIST_MODEL_URL, and EXIST_MODEL_CHAT). Your ~/.config/opencode/opencode.json is merged over it.

What you type into claude and opencode is captured to workspace/.capture/ (secrets redacted, never responses). CODE_SERVER_CAPTURE=false turns it off. See Capture.

No telemetry. code-server runs with --disable-telemetry, Claude Code with DISABLE_TELEMETRY and DISABLE_ERROR_REPORTING, opencode with sharing off. Your conversation goes to the AI provider you use, under its terms.

pi is pi, baked into the image and pointed at the same chat model on every boot, with the MCP servers of every add-on that is on (qmd, Firecrawl, Playwright):

pi                                            # the TUI
pi -p "what is still open in notes/plan.md?"  # one answer

It is the same CLI and the same configuration (workspace/.decree/lib/pi.sh) decree's agent_task runs. /login to a hosted provider if you want one too; it is kept beside the stack's own.

Installs and logins persist on the code_server_data volume: npm's global prefix, ~/.config, ~/.local, ~/.cache, Claude Code's config and pi's all live there.

A tool that keeps state elsewhere (~/.mytool/, ~/mytool.json) needs one step before you log in:

docker exec code-server mkdir -p /code-server-data/home/.mytool
docker exec code-server touch   /code-server-data/home/mytool.json

Anything in /code-server-data/home/ is linked into $HOME on every boot.

This container mounts workspace/, not the repo. Edit the stack on the host.

Notes​

  • Runs on the shared existential/decree:local image, with the same toolchain as the automation containers.
  • code_server_data (installs, extensions, logins) isn't backed up; losing it means reinstalling and logging in again.
  • One shared password, CODE_SERVER_PASSWORD. Its shell can read and write all of workspace/ and reach every model and MCP server, so treat it as root-equivalent.