Create a new Claude Code cloud session via claude --cloud on a pseudo-terminal (the CLI rejects this form when stdout is a pipe). The work runs as a background job and this call returns within a few seconds. If status is done, the result is final: the parsed session_id and session_url, or, when parsing fails, parsed false and the CLI output (terminal controls removed) in cli. Otherwise poll claude_get_session with job_id until job.status is done or the call returns an error; failures surface there with the same error codes. A github.com repo is cloned into CLAUDE_MCP_REPO_CACHE (origin set to that GitHub repository) and, when base_branch is set, checked out there before the CLI runs. Any other absolute directory must be on CLAUDE_MCP_REPOS. Omit repo to use CLAUDE_MCP_CWD. When Claude Code asks to trust the chosen folder, the call fails with WORKSPACE_NOT_TRUSTED unless CLAUDE_MCP_TRUST_CWD=1. The cloud VM's clone, push, and pull request use the Claude GitHub app (or /web-setup), not this server. If the session later can't push its branch to GitHub, ask it via claude_send_followup for a git format-patch (or full git diff) of its changes and apply, test, push and open the PR yourself.

Queue a follow-up message into an existing Claude Code cloud session via claude --cloud <session> -p <message>. Replies are only visible at the session_url; claude_get_session reports the pushed branch and pull request. If the session can't push to GitHub, use this to ask for git format-patch origin/<base>..HEAD --stdout (or a full git diff including new files), then apply it locally with git am / git apply --3way and push it yourself.

With a job id from claude_create_cloud_session, report that create's progress in job (status, elapsed time, the tail of the CLI output, and session_id once the CLI prints it), its final result once job.status is done, or its error. With a job id from claude_run_local, report that run in local_job the same way. A job read uses memory only and runs nothing. With a session id or URL, report the branch a cloud session pushed and any pull request it opened. Prepares the github.com checkout in CLAUDE_MCP_REPO_CACHE (origin set to that repository), then runs print-mode claude --teleport=<session> --debug there and stops when the CLI exits. Print mode checks out the session branch and exits; it is not given a prompt and it is not started on a terminal, so it does not continue as a local interactive session. The branch is taken from the debug line Branch after checkout when that matches HEAD, otherwise from a HEAD move. The pull request number and URL come from that output when present, otherwise from a read-only gh pr list. Claude's GitHub app (or /web-setup) is what pushes the branch and opens the pull request; this tool only reads them.

Stop a claude_create_cloud_session job or claude_run_local job that is still running: every process it started on this machine (git, the pty helper, claude --cloud or claude -p) gets SIGTERM, then SIGKILL 5 s later. Pass the job_id, or for a create job the session id or URL it already printed. Waits up to about 6 s for the processes to exit and reports stopped. This does not stop a cloud session: one the CLI already created keeps running at its session_url. A job that already finished is left as it is (cancelled false), and one that failed returns its error.

Run Claude Code headless on this machine (claude -p) in an existing checkout, instead of a cloud session: for work that needs the local toolchain or hardware. cwd must be inside a CLAUDE_MCP_LOCAL_ROOTS directory after resolving symlinks; with no roots configured every call fails with LOCAL_RUNS_DISABLED. Unless CLAUDE_MCP_LOCAL_ALLOW_SHELL is set, the run uses Claude Code's --restricted mode: no shell or code tools, no WebFetch, no MCP servers, and the checkout's settings, hooks and .mcp.json are not loaded (needs Claude Code 2.1.248 or later). The run is a background job and this call returns within a few seconds: if status is not done, poll claude_get_session with job_id until local_job.status is done or the call returns an error. The final result has the session_id (pass it as resume to continue), the result text, turns, cost and usage, denied tools, and the files git status lists. permission_mode defaults to dontAsk, which denies anything not read-only or in allowed_tools; permission bypass is never used. Pre-approving Bash or using auto mode runs shell commands unasked and fails with LOCAL_SHELL_DISABLED unless CLAUDE_MCP_LOCAL_ALLOW_SHELL is set. claude_cancel_session stops a run.

Report how much of the Claude plan's usage allowance is used: each window (the rolling session window, weekly windows, model- or surface-scoped ones) with its percent, reset time and severity, plus usage-credit spend. Runs print-mode claude -p --restricted -- /usage, a local command that makes no model call and costs nothing; readings are reused for 60 s unless refresh is true. When CLAUDE_MCP_USAGE_STOP_PERCENT is set, blocked_by names the window that makes create and claude_run_local fail with USAGE_LIMIT_NEAR. windows is null when the CLI can't read plan usage.

Report whether the local Claude Code CLI is signed in, via claude auth status. If it is not, sign in with claude first (run it once in a terminal and complete the claude.ai login). A missing CLI returns CLI_NOT_FOUND with the install command (npm install -g @anthropic-ai/claude-code).