Claude in Deploy Center

Deploy Center can ask Claude three questions about your deploys, with your own Anthropic API key. It is opt-in and off by default.

Question	Where	What is sent
Explain this failure	detail sheet and build log, when the last build failed: Ask Claude	the part of the build log around the first error, masked
Release notes for humans (EN/DE)	detail sheet, when commits are waiting: Ask Claude	the titles of the waiting commits and PRs (no authors, no ids)
Is this safe to ship?	detail sheet, when commits are waiting: Ask Claude	the diff stats, the risky file names (migrations, env files, lockfiles, build config, workflows), the commit titles, whether auto-deploy is on

Explain this failure also offers Fix with Claude, the same button as on the row: the failed build goes to your Autofix routine (deploys autofix --only <platform> <project>), under the usual caps. Without Autofix set up, it offers the setup instead.

Turning it on#

Settings → Claude in the app → Ask Claude with my own API key. Paste a key from console.anthropic.com, Save, then Verify (it looks up the model with the key; that costs nothing). The Ask Claude menu shows up only when the switch is on and a key is saved.

Model: Claude Opus 5.5 by default; Claude Sonnet 5.5 and Claude Haiku 5.5 are cheaper.
Monthly cap: 300,000 tokens by default. It is checked before every request; once reached, nothing is sent until the next month or until you raise it. The tally is kept on this Mac (~/.config/deploys/ai-usage.json), so the Anthropic Console stays the source of truth for billing.

Privacy#

No server of ours and no telemetry: the app (through deploys) talks straight to api.anthropic.com.
The key lives in the macOS keychain (item deploy-center-anthropic), written by deploys ai set-key over stdin. It is never printed and never passed as a command argument. ANTHROPIC_API_KEY in the environment is used instead when it is set.
Everything sent goes through the same masker as Autofix first: API keys and tokens (sk-…, ghp_…, github_pat_…, xox…, AKIA…, rnd_…, JWTs), token= / password: style values, passwords in URLs, and email addresses. A help, not a guarantee: read SECURITY.md.
The prompts tell Claude to treat logs and commit titles as data, never as instructions.

Command line#

The app runs these; you can too.

pbpaste | deploys ai set-key          # store the key (or set ANTHROPIC_API_KEY)
deploys ai verify                     # the key works and the model is available
deploys ai explain my-project         # why the newest build failed, and the likely fix
deploys ai notes my-project --lang de # plain release notes, English (default) or German
deploys ai ship my-project            # Risk: low, medium or high, and what to check
deploys ai model claude-sonnet-5-5    # claude-opus-5-5 (default), claude-sonnet-5-5, claude-haiku-5-5
deploys ai cap 500000                 # tokens per month
deploys ai status                     # key set (never the key), model, cap, used this month
deploys ai forget-key

Add --json to explain, notes and ship for {text, model, tokens, used, cap}.

The request is the Messages API (anthropic-version: 2026-10-21) at low effort. On Opus and Sonnet the server-side fallback is on (fallbacks: "default"): if a safety classifier declines, the API answers with a fallback model in the same call. A refusal that remains is reported as such.

"What this ships" on every pull request#

examples/claude-pr-ship-check.yml is a GitHub workflow for your app's repo. On each pull request it:

runs deploys changed --json through the Deploy Center action (report only, it never deploys),
lists every project that deploys from this repo with its auto-deploy setting (deploys vercel json / render json) and what already waits on it,
runs Anthropic's Claude Code GitHub Action (anthropics/claude-code-action v1, pinned to a commit) with that list in the prompt; Claude reads the PR diff and posts one comment, What this ships: the projects affected, a risk level, and whether auto-deploy is off (then nothing ships on merge until someone deploys).

Claude may only run gh pr diff, gh pr view and gh pr comment. It needs the secrets ANTHROPIC_API_KEY and VERCEL_TOKEN and/or RENDER_API_KEY, and the workflow permissions contents: read, pull-requests: write, id-token: write. Pull requests from forks get no secrets and are skipped; do not switch it to pull_request_target.

Other ways to use Claude with Deploy Center#

Copy for Claude in the build log: a ready prompt with the masked log, for any Claude chat.
Autofix: a failed build goes to a Claude Code routine that opens a fix PR.
Settings → Claude connects Claude Desktop and Claude Code (the MCP server and the Claude Code plugin).