Anthropic’s Claude Code release of 17 November adds a way to register an MCP server without putting its token on the command line. claude mcp add-json <name> - reads the server’s JSON from stdin. The same release asks for confirmation before /install-github-app changes a repository, and adds a /status list of broken configuration.
Tokens off the command line
The release notes describe the new form as reading “the server’s JSON from stdin, so tokens can stay off the command line”. It isn’t available inside a Claude Code session, so you run it from your own terminal. The dash at the end tells the command to read stdin. The sketch below is mine, not Anthropic’s. The URL and the EXAMPLE_TOKEN variable are placeholders.
printf '{"type":"http","url":"https://mcp.example.com/mcp","headers":{"Authorization":"Bearer %s"}}' "$EXAMPLE_TOKEN" | claude mcp add-json example -

Safer defaults
Several changes make Claude Code ask more often, or refuse more, where a mistake costs the most.
- Repository changes:
/install-github-appnow asks before it changes a repository. Its repository, secret and API key steps also use the standard list and keys, with shorter messages, a spinner and clearer errors. - Secrets in commits: on newer models and in PowerShell, Claude is now told not to commit files that likely contain secrets.
- Auto mode: it now ignores more allow rules that run any package, such as
Bash(pnpm dlx *)and path or version forms like/usr/bin/npm run:*. The safety check decides those instead. Its notice about falling back to billed classifier requests now says an upcoming release removes the fallback, and each action it would have checked will be asked about or denied. - Paths: path checks refuse some symbolic links, and some paths that used to count as inside a working directory now ask for permission.
- Outside reads:
permissions.blockReadsOutsideWorkingDirectoriesasks about more Bash commands in bypass and auto modes, denies some in auto mode, and refuses more commands that name a path under a Read or Edit deny rule. - Long variables: Bash commands whose variables expand to very long text are denied, with a message to use shorter values or a file.
- PowerShell and Chrome: the PowerShell tool asks before one more command, and Claude in Chrome refuses to upload some credential files.
One change cuts the other way. To speed up a session’s first Bash command, shell start-up files may now run before the command is approved, even if you then refuse it. Set CLAUDE_CODE_DISABLE_SHELL_HEAD_START=1 to turn that off.
Fixes that stop files being lost
Five changes cover ways Claude Code could delete or overwrite something you wanted to keep.
.mcp.json:claude mcp addandclaude mcp removeat project scope no longer delete other top-level keys, such as$schemaor another tool’sservers.claude mcp add --scope projectnow stops, instead of overwriting the file, when.mcp.jsonlists its servers at the top level with nomcpServerskey./export: it no longer overwrites an existing file without asking. The dialog asks, and/export <name>keeps the file and tells you to run/export.- Remote Control: stopping the server in worktree mode no longer deletes a running session’s git-ignored files, such as
.env. - Self-hosted runners: with
--use-anthropic-git-proxy, a runner no longer deletes a non-empty directory named byGIT_CONFIG_GLOBALat startup. It keeps the directory, ignores the variable and warns. - Worktrees:
ExitWorktreenow asks before removing a worktree that may hold unsaved work.

A health list in /status
/status now lists five kinds of problem. /config adds a note when other settings override a row. A settings change from /config that fails to save now puts a line under the prompt saying so, and /rewind says when some files couldn’t be restored.

Resuming sessions loses less
A resumed session now shows a line when its earlier messages can’t be loaded. A transcript read that ends early is retried once. Headless sessions no longer lose stored messages on resume when a message was sent again with the same UUID after the process was killed. Tasks created after /clear are no longer lost in cloud and self-hosted runner sessions, and a compaction right after a reply cut at the output token limit no longer drops that reply’s tool call and result. Subagents launched without a model now stay on the model they started on when resumed, including after /model.
Gateways, billing and Code Mode
Teams running Claude apps gateways get two additions. The attributionProject setting names a session’s billing group, for gateways that set admin.billing_group_tag. Gateway hint headers now carry x-claude-code-managed-settings-checksum, sent where managed settings set CLAUDE_CODE_GATEWAY_HINT_HEADERS=1. /tui classic is now the name for the non-fullscreen renderer, and /tui default still works.
Code Mode, which we covered in our report on Claude Code’s Code Mode, no longer refuses to start when an installed mod hooks tool calls. An administrator can still keep it off in managed settings. A running Code Mode script in auto mode also stops getting permission prompts and denials once Claude starts a second script that auto mode allows.
Smaller changes worth knowing
| Area | What changed |
|---|---|
| WebFetch | Refuses a URL over 2,000 characters before any permission prompt |
@ mentions | Only the first 100 mentions in a message are looked up as files |
| Remote MCP servers | Most servers reached by hostname over https are retried in the background for about a day after an outage such as a VPN drop |
| Grep | Content and count results are sorted by file path, which fixes repeated or skipped results when paging with head_limit and offset |
claude -p | Goes on without a model’s settings after 10 seconds if the fetch stalls, such as at a proxy |
| Windows | Faster start-up with no Git Bash and no pwsh on PATH, and PowerShell commands run up to about half a second faster |
| VS Code | New “Claude Code: Add to Chat” in the Explorer’s right-click menu, which @-mentions the selected files and folders |
The full list in the release notes runs to more than a hundred lines, most of them terminal, hook and agent fixes.
Discover more from Sulat AI & Tech News
Subscribe to get the latest posts sent to your email.




