Agents inside a workspace

Claude Code and Codex run headless on the machine, one process per turn, launched detached so a nap mid-turn stalls the reading and not the agent. This page is how each is run, what the machine tells them, and where they differ.

How a turn is launched

The runtime writes a small script to the machine and starts it in its own session, setsid bash <script> > <log> 2>&1 &, then polls the log every second and a half. The process belongs to the machine, not to the host: a host restart re-opens every run the machines still hold and reads the rest of the output. Every command starts placed after the helper for the memory killer, so a build that outgrows the machine loses itself and not the link.

Claude Code runs claude -p with stream-json in and out, --verbose, the access flag, --model, --effort, --name <title> when a title was given, and either --session-id <uuid> for a new session or --resume <uuid> for a send. The task is never an argument; it rides stdin as one JSON line, and stdin closes when the reply lands, which ends the process after its current turn. The session id is minted before launch so the thread is addressable before the CLI prints anything. Its config lives at /root/.claude-cfg through CLAUDE_CONFIG_DIR.

Codex runs codex exec with --json, --skip-git-repo-check, resume <thread id> for a send, the sandbox and approval flags, -m <model> and -c model_reasoning_effort=<effort>. The task travels in a quoted heredoc on stdin, because a long task as an argument would hit the kernel's per-argument cap. Its home is /root/.codex.

What the machine tells every agent

Every installed agent gets the machine's context through its own always-loaded hook, /etc/claude-code/CLAUDE.md for Claude Code, /etc/codex/requirements.toml for Codex, and the full document as a skill named wsp-machine. Both open with the line that wsp writes the file when a workspace is forked and edits are overwritten. The facts in it:

• Containers do not run here; services install natively.
• wsp-daemon listens on 0.0.0.0:7070 with its own token and must not be stopped or shadowed.
• A background process started with a plain & inside a tool call dies when that call ends; a server that must outlive the turn starts with setsid nohup ... &.
• A turn ends when the agent process exits, not at its reply, and a reply given with a command still running reads failed.
• A loopback-only bind is unreachable through the preview edge; bind 0.0.0.0. Ports below 1024 are not forwarded.
• Secrets are exported by /etc/profile.d/wsp-secrets.sh, used by name, and that file is never read, printed, logged or committed.
• .env files, .netrc and private keys were never copied from the person's computer.
• Sign-ins go through wsp: BROWSER is the shim, the page opens on the person's computer, and the callback is tunnelled back.

Your MCP servers travel with them

The MCP servers you configured for an agent go onto the image inside that agent's own config file, ~/.claude.json for Claude Code and ~/.codex/config.toml for Codex. Laptop paths are rewritten to the machine's, a command sitting under a known bin directory becomes its bare name, and a server that carries a secret travels only when you answered copy for it. Each is recorded as installed, fetched on first use for npx and uv commands, or skipped with the reason.

Titles and renames

A few seconds after the first turn announces its session, wsp asks the agent's cheapest model for a name, three to six words, and reads what the agent already calls the session. A name a person gives is never replaced. A rename writes into the agent's own store: Claude Code's session file gets the same record its own /rename appends; Codex's SQLite state gets its name column set. So the agent's own session list and wsp show the same name.

Where they differ

	Claude Code	Codex
message into a running turn	joins it	waits behind it
cost reported	yes	no; duration only
context window	200k or 1m	none
default access	bypassPermissions	danger-full-access
sign in on the machine	claude auth login	codex login --device-auth
install size	about 208 MB	about 455 MB
session store	projects/<folder>/<id>.jsonl	state_<n>.sqlite

Codex needs sqlite3 on the machine for titles and renames; it is on the floor of every golden. Neither agent can be named before it announces a session.

Signing them in

Claude Code takes an API key at create time, ANTHROPIC_API_KEY from your .env, or a subscription sign-in on the machine during wsp init. Codex signs in on the machine the same way, or reads OPENAI_API_KEY from the machine's environment. The OAuth credential for Claude Code is never copied from your computer; see Sign-ins and the relay.