Claude Code 2.1.259 adds headless deny-mode and org-managed MCP
Latest release retools Claude Code for automation and teams: GitLab MRs now surface like GitHub PRs, and parallel sessions stop clobbering each other's config.
Copy markdown
Headless hosts can auto-deny every prompt
New `--permission-prompts none` denies anything that would pop an approval dialog, while the active permission mode (auto included) keeps deciding on its own. It's the missing piece for running Claude Code in CI, cron, or a container without it hanging on a prompt.
Orgs can push MCP servers to everyone
A new `managedMcpServers` setting lets admins provision HTTP/SSE MCP servers to every user, same shape as `.mcp.json` (entries that name a command to run are skipped). Standardize your team's tools centrally instead of asking each dev to edit their config.
GitLab MRs get GitHub-PR parity
Claude Code now recognizes `glab mr create/merge/close/reopen/note/update`, rendering merge requests as `MR !N` in the collapsed tool summary and refreshing the footer MR badge. GitLab shops finally get the inline MR view GitHub users already had.
Parallel sessions stop overwriting your config
Fixed sessions silently reverting each other's `~/.claude.json`, so workspace trust and MCP/project state are no longer lost when you run several Claude Code sessions at once. Keep two terminals open and your settings finally stick.
`claude plugin validate` speaks JSON
Plugin validation now takes a `--json` flag for a machine-readable report, so you can gate plugin publishing inside CI instead of eyeballing console output.
Elsewhere: 207 WebGPU kernels for browser-local AI
Hugging Face shipped `@huggingface/kernels`, a JS library that pulls optimized WebGPU kernels straight from the Hub — 2.57x faster than ORT WebGPU on an M4 by geomean across 809 cases. `npm install @huggingface/kernels@preview` and run inference client-side, no server.