GoLive

Take your agent-built product live: hosting, database, auth, domain, email, payments — on your own accounts. Then hand it over, or tear it all down.

Your coding agent can build an app in minutes. Getting it to real users still means accounts, hosting, databases, domains, secrets and connected services. GoLive is the open-source Agent Skill for that work: it detects what your app needs, plans the exact changes, asks for your approval, applies them with your own logins, and verifies what actually works — then records what it created, re-checks it for drift on demand, and can remove it again.

Automate the parts providers expose. Guide you through the parts that need a human. Verify what can be observed, and make unfinished work clear. No GoLive account, hosted backend or product telemetry.

Early alpha · 0.1.0-alpha.3 Disposable live tests now cover six journeys: hosting (Vercel, Netlify), database (Supabase, Neon), custom-domain DNS (Porkbun, GoDaddy), transactional email (Resend), test-mode payments (Stripe) and Supabase authentication, plus the teardown uninstall path. The ownership document and the on-demand golive status drift check are implemented with test coverage (golive status also ran read-only in a live validation), while the broader roadmap is our direction, not a claim that it is all built.

Whether to give an agent your provider accounts comes down to four questions. These are this project's answers, with the limits stated where they exist.

• You still approve every write. Nothing reaches a real account without a plan you have seen and approved: apply refuses without that plan's id and --yes, and it re-checks the plan's identity before writing, so a changed release or config invalidates the old approval. DNS writes need --confirm-dns, deletions need --confirm-destroy, and live-mode steps — live payments, production data, a real account — need --confirm-live, which now includes a project's first production deploy, because approving a plan alone used to be enough to write production for the first time. Credential values are read only in-process, never printed, and never in arguments, plans, state or reports; the file golive stores them in is plaintext at mode 0600 outside your repo, not a keychain. One limit worth naming: those flags are arguments the agent passes on your behalf, and an agent already logged in to your provider can write there with no golive plan at all. Trust, access and control separates what the code enforces from what is only an instruction the agent is asked to follow.
• A run stops rather than pushing on. apply stops at the first failed check, missing confirmation, missing prerequisite or provider that contradicts the plan. Later steps do not run, and the next apply resumes at that step. Recovery covers reading the failure, which steps resume, and the cases that need a reviewed decision first.
• Rollback is narrow, opt-in and never automatic. A failed check never triggers a rollback. release.rollback: true plans one step that re-points production at an earlier deployment golive itself recorded; a deployment built by a dashboard, a Git push or a pull request is not a target, and it touches no data, DNS, payment or email resource. Only Netlify supports these re-points today — on Vercel you correct production in the dashboard (Vercel's adapter has no read of what production serves). Promotion and rollback are implemented and mock-covered, not live-validated.
• Nothing is left behind silently — which is not the same as nothing being left behind. golive teardown removes only resources it can prove it created, re-reads the DNS zone and the host project after deleting, and names every leftover it cannot remove — Supabase and Neon projects, the Resend sending domain, a zone or host project it cannot read — as a handoff saying what remains and how to remove it by hand. A removal also forgets the baseline golive recorded for that resource, so golive status does not report golive's own teardown as drift.

Those answers in full: trust, access and control and recovery. The architecture is the product contract, provider scope says what each provider can do today, the validation record separates what has been exercised live from what is only mock-covered, and distribution covers installation and updates.

Install · Use GoLive · See the workflow · Alpha scope · Roadmap · Contribute

You need Node.js 20+, npm/npx, Git, and a coding agent that can load skills and run commands. Installation has been checked for Codex and Claude Code; other clients are unverified.

Install once for all your projects. Run this from any directory:

npx skills add https://github.com/mikehasa/golive-skill --skill golive --global

Select your agent when prompted: use the arrow keys to move, Space to select, and Enter to confirm. That screen is waiting for input; installation continues after you confirm.

To skip the agent picker, use the command for your agent:

# Codex
npx skills add https://github.com/mikehasa/golive-skill --skill golive --global --agent codex --yes

# Claude Code
npx skills add https://github.com/mikehasa/golive-skill --skill golive --global --agent claude-code --yes

For installation in just one project, run from that project's repository and omit --global.

Or paste this into your coding agent:

Install the GoLive skill globally so I can use it across projects:
npx skills add https://github.com/mikehasa/golive-skill --skill golive --global

Target the agent I'm using: add --agent codex --yes for Codex, or
--agent claude-code --yes for Claude Code. Keep --global.
If the agent isn't clear, ask me which one.

Verify the installation with:
node <installed-skill-dir>/scripts/golive.mjs version --json
Tell me if I need to reload skills or start a new session.
Stop after installation; don't connect accounts or deploy yet.

The install includes the instructions, provider references and prebuilt runtime. It does not connect accounts or deploy anything. See installation and updates for noninteractive agent flags, runtime verification and the optional own installer.

The same skill is published to npm as golive@0.1.0-alpha.3 (dist-tags alpha and latest), which installs it offline, with no Git or Skills CLI involved:

# Codex
npx golive@alpha install --agent codex

# Claude Code
npx golive@alpha install --agent claude

Add --global to install into your home directory (~/.agents/skills/golive or ~/.claude/skills/golive) instead of the current project. The installer copies the complete skill the package ships with, refuses an existing destination, and never connects provider accounts.

Both channels carry the same release. The npm package publishes the version in this repository, including the standalone installer helpers, so an npm installation is an owned copy that updates in place. The earlier 0.1.0-alpha.0 snapshot has no updater: remove that copy and reinstall, or use the GitHub channel, which manages its own installs. The npm package also exposes the terminal CLI: npx golive@alpha help, version, detect, menu, plan, apply, verify, handoff — apply needs the approved plan ID and explicit confirmation. See installation and updates for the channel's exact limits.

Open your app repository in your coding agent. After installation, reload skills or start a new session if GoLive does not appear. The skill's name is golive. For the standalone installation above, type one of these in your agent's chat:

You can also ask in plain language:

Use the golive skill to take this app live. Keep the providers it already uses.
Show me the destination accounts and plan before changing anything.

GoLive inspects the app, asks about missing providers, and presents the destination accounts and plan for your approval before making provider changes.

These are chat prompts. golive skill is not a command. Installing through npx skills add does not add a golive command to your terminal; the agent runs the included CLI from the installed skill directory. The npm package can also install the skill offline and exposes that CLI as npx golive@alpha <command>; its terminal commands run individual operations rather than the conversational skill workflow.

An illustrative conversation for an app that already uses Supabase:

You: Use golive to take this app live. Use test resources for now.

Agent: This app uses Supabase. Where would you like to host it?

1. Vercel — built-in hosting workflow; live-tested with Supabase.
2. Netlify — built-in hosting workflow; the live-tested pairing is Netlify + Neon.
3. Other — tell me your provider; I'll try a guided setup using its official tools and docs.

You: Vercel.

Agent: We'll keep Supabase. Does this app already have a Supabase project, or should we create a new test project?

You: Create a new one.

Agent: Run vercel login and supabase login in a separate terminal window. I'll check which account and organization are connected when you're done.

You: Done.

Agent: Here's the plan: create a new test frontend project in your selected Vercel team, create its database in your selected Supabase organization, transfer the required environment values, and deploy. I'll show the actual destination names and IDs, resource settings and cost information with the plan. Approve before I apply it.

You: Approved.

Agent: The deployment is live. Here is the URL, the verification report, and anything still unverified. Next we'll test the app's own signup and data flows. When you're done with this test project, golive teardown plans its removal for your approval before anything is deleted.

GoLive keeps the provider choices already in your app and asks about missing pieces. You handle signups, browser logins, identity checks and purchases. If an API key is needed on macOS, a native hidden-input dialog explains why it is asking and where the key will be saved. Its value goes directly to the local credentials file, never to chat or command output. Other platforms use your own editor as a fallback. A later change to the plan needs another approval; connecting auth or a domain may require a follow-up after the first deploy.

Using another provider? The skill has a general guided flow: check the provider's official CLI, an available official MCP integration or API, then guide you through its dashboard if needed. The agent still shows the destination, changes and cost before asking for approval, and checks what it can afterward. This is best-effort guidance, with no guarantee of completion or the same verification coverage as a built-in adapter. If a step cannot be completed or verified, you get the specific blocker and next action. See guided provider scope.

Two hosting choices: Vercel and Netlify. Two database choices: Supabase and Neon.

These were approved disposable runs on existing accounts; completed test resources were deleted afterward, and recent runs' disposable projects and records are cleaned up under the same supervision. Cross-pairings have mocked coverage, not equivalent live proof. Supabase CLI-login reuse separately passed read-only verification; the complete deployment test used an explicit token. A new user's first-account setup and every application framework have not been validated.

The lifecycle commands have their own evidence: golive teardown was live-exercised on a disposable Netlify project (blocked without --confirm-destroy, then removed, the account's site list unchanged apart from it) and earlier runs removed the GoDaddy and Porkbun records golive had written, revoked the Resend sending keys it had issued and removed the Stripe test-mode endpoint it had registered. golive status ran read-only against a live project; golive handoff --write ran on a disposable Vercel-only fixture: both artifacts were written and audited (a provenance tag on every claim row, the ownership proof and the teardown gate named, no credential-shaped value in the document, its JSON twin, state or config), and the project was removed afterwards through the approved teardown flow — the stack was host-only, so the document's other-provider rows remain mock-covered. See observed validation for the evidence.

Experimental adapters also exist for Supabase Auth configuration, the Supabase Auth signup journey, its password recovery and account isolation, and Cloudflare DNS. Supabase Auth settings — signup, email confirmation, minimum password length, the mailer it uses, plus the site URL and redirect allowlist — are automated through an approved plan and re-read for evidence, and that path passed a disposable live run: the policy write held in the read-back (password minimum length: 6 → 12) and auth-policy ended with the built-in-mailer advisory as its only finding. The opt-in signup journey (auth.e2e) passed the same run: one approved step seeded a real test account (auth:test-user, needs --confirm-live), the address could not sign in before confirming (email_not_confirmed), and the auth-signup/auth-session checks proved the signup email, the enforced confirmation, the confirmed login, the session token and the anonymous refusal. Two limits stay: the confirmation was applied through the Auth admin API rather than the seeded account's own email click, and inbox delivery is human-confirmed by design — golive never sees the inbox. A later approved run on a disposable fixture (a deployed Vercel site whose declared route answers 401 without a session, plus one RLS-protected table) exercised both app-side legs: an anonymous GET of auth.protectedPath answered 401 and the signed-in probe read that table as the authenticated user, so the probe's bearer fix is no longer mock-covered. What that evidence cannot show: that run's table line is a count rather than table names, and any 401 counted as protected — a WAF or maintenance page would read the same; both were fixed afterwards (issue #30: the probe names the tables it read, and a refused protected path is corroborated against the public root, with mocked coverage and no live re-run yet). Password recovery is live-validated on the same provider: the same 2026-09-24 run carried auth.smtp: resend (the custom-SMTP write and the raised auth email rate limit, both read back) and auth.recovery: true, whose one approved step (auth:recovery, needs --confirm-live) rotated that recorded test account's password through the provider's own recovery calls — request the email, mint the link with the admin API, exchange it for a session, set the new password with that session — and the auth-recovery check passed every leg: the request was accepted, an address with no account got the same answer (no account enumeration), the spent token was refused on replay, the new password signed in and the one it replaced did not. The inbox click and any captcha stay with the human (the auth:recovery-email handoff says so), the SMTP password is write-only (the provider answers a hash, so the read-back proves the settings, not a delivery), and the account was confirmed through the Auth admin API rather than the owner's click. Account isolation is implemented on the same provider too: auth.isolation: true with auth.identityPath and auth.isolationPath adds one approved step (auth:isolation, needs --confirm-live) that seeds a second real test account — the address derived from auth.testEmail, the password again only in that run's memory — and confirms it through the provider's admin API (no second inbox click: the journey is about the app's data, not delivery). The auth-isolation check then signs in as both accounts and reads the app's own two declared routes on the production URL: both must refuse an anonymous caller (a 200 is a critical finding), each account's identity route must answer with its own user id and never the other's, and the rows route must return only the caller's own rows — checked with one unique marker row per account written through that route with the account's session and read back, so another account's marker in the answer is a cross-account read and fails critically. When the routes are not declared, the non-blocking auth:isolation-routes handoff hands the app-code task over; a 404 or a refused session skips with that task named, never as a pass. Account isolation is implemented and mock-covered, not live-validated yet — its live run comes separately. The recovery run's own output contained two defects, both fixed here with mocked regressions: teardown reported the owner's adopted sending domain as created by golive (an empty creation-marker list made [].every() true, so every recorded domain read as golive's), and the auth:recovery-email handoff showed a standalone verify skip as its evidence while state recorded that step done. Its third finding — Resend kept reporting that domain verified while the records it listed were absent from the zone's authoritative nameserver — is fixed by #52: email-verified now resolves the records the provider itself lists for the domain before passing (a verified domain whose records are gone fails, a record golive wrote inside the propagation window only warns, and a provider that cannot list them skips rather than passing), and the email plan keeps the email:dns step or handoff for records that do not resolve, so a stale flag can no longer hide them. Mock-covered; not re-exercised live. The DNS, email and test-mode payment paths listed above are the tested ones, with the custom-domain runs using Porkbun and GoDaddy record writes; Cloudflare DNS specifically is not a validated alpha path yet, and other auth providers stay guided. See provider scope and observed validation.

A working URL is the beginning. Depending on the app, going live can mean all of the following. GoLive should work out which items apply, help you finish them, and show evidence for the result. A static site should not be asked to set up a database; a paid SaaS should not stop at a deployed homepage.

This is our product roadmap as a launch checklist. Checkmarks and strikethroughs mark specific live-tested milestones, not a finished category or a completed checklist for your app.

✅ Live-tested · 🚧 In progress / experimental (code exists; complete journey pending) · 🗺️ Planned

• ✅ Frontend hosting: Prove deployment on Vercel and Netlify. Build, deploy and verify the intended project on the two tested paths.
• ✅ Database: Prove provisioning and connection with Supabase and Neon. The tested paths include environment wiring and application CRUD checks.
• ✅ Environment wiring: Connect hosting and database credentials on both tested paths. Broader secret rotation and environment lifecycle management remain planned.
• 🗺️ Backend / servers: dedicated API services, containers, persistent servers, runtime configuration and health checks. App routes already deploy through the supported hosts.
• 🗺️ Schema and data: reviewed migrations, safe rollout, environment separation and app data checks. These were separately supervised in live tests; a reusable workflow is still planned.
• 🗺️ File and object storage: buckets, uploads, access rules, signed URLs and lifecycle policies.

• 🚧 Authentication: signup, login, sessions, password recovery and account isolation. Supabase auth policy (signup, email confirmation, minimum password length, mailer) and the site URL/redirect allowlist are written through an approved plan, re-read for evidence and verified by the auth-policy/auth-redirects checks — exercised in an approved disposable run, where the policy write held at a twelve-character minimum. The opt-in journey (auth.e2e: true) passed the same run: the auth:test-user step seeded a real test account, that address could not sign in before confirming, and the auth-signup/auth-session checks proved the signup email, the enforced confirmation, the confirmed login and the session token — live-validated for Supabase across disposable projects, where the confirmation came through the Auth admin API instead of the seeded email click, inbox delivery stayed human-confirmed, and a later run proved a declared protected path (an anonymous 401) and a signed-in read of an RLS-protected table, reported as a count rather than a table name. Password recovery is live-validated on the same provider (auth.recovery: true adds the auth:recovery step and the auth-recovery check, which proved no account enumeration, a one-time token and the replaced password on a disposable project; the confirmation came through the Auth admin API, the inbox click stays human-confirmed, and two output defects that run found — a false "created by golive" ownership claim and a handoff evidence text contradicting the recorded step — are fixed with mocked regressions). Account isolation — the other half, and the one earlier runs could not exercise — is implemented and mock-covered the same way: auth.isolation: true with auth.identityPath and auth.isolationPath adds the auth:isolation step (a second real test account, confirmed through the provider's admin API and recorded by id and address) and the auth-isolation check, which signs in as both accounts and proves on the app's own routes that neither can read the other's identity or rows (a cross-account read fails critically; an undeclared or 404 route skips with the app-code task). Its live run comes separately too. Other auth providers stay guided.
• 🗺️ OAuth / social login / SSO: client registration, consent screens, scopes, callback URLs and provider reviews. Current auth-provider setup is guided.
• ✅ Payments and subscriptions: Prove test-mode checkout and webhook acceptance with Stripe. A real test-card payment delivered a signature-verified checkout.session.completed event. Live-mode readiness, entitlements, refunds and subscription events still need validation.
• ✅ Transactional email: Prove sending-domain setup, verification and real delivery with Resend. A send through the app's own environment key was delivered (to spam on a fresh subdomain, no DMARC yet). With auth.smtp: resend the auth:smtp step also writes the auth project's custom SMTP — Resend's host/port/user, the sender the app already uses, and an SMTP password taken from a sending key golive issued (the email journey's key, or one it issues for SMTP alone) — and raises the project's own auth email rate limit (rate_limit_email_sent) to 30 per hour (or auth.emailRateLimitPerHour) in the same write, because the provider keeps that limit with custom SMTP. auth-policy then reports custom SMTP via Resend instead of warning about the built-in mailer. The password is write-only (the provider answers a hash), so the read-back confirms the settings and a real auth email is the only full proof. Live-validated on a disposable project (2026-09-24): the same run wrote the custom SMTP and read it back (smtp.resend.com, port 465, user resend, sender auth@mail.trytofu.xyz) together with auth email rate limit: 2 → 30 per hour, issued the SMTP key by itself and revoked it in teardown, and auth-policy then read custom SMTP via Resend with 30 auth emails/hour — settings and rate limit only, since the password itself can never be read back. Bounce handling, richer message content and actual inbox delivery (human-confirmed by design, and doubtful on that run's domain — see issue #52) still need validation.
• ✅ Domains / DNS / HTTPS: Prove domain attachment, DNS wiring and HTTPS serving on host+DNS pairs. Tested: Vercel attachment with Porkbun and GoDaddy record writes under --confirm-dns, ownership verification and HTTPS 200 on disposable subdomains. The Cloudflare DNS adapter, redirects and further host pairings still need live validation.
• 🗺️ SMS and push notifications: sender registration, credentials, permissions and delivery checks.
• 🗺️ Third-party and AI services: API access, scopes, callbacks, quotas and functional tests. Missing environment variables are detected today; service-specific workflows are planned.
• 🗺️ Background work: cron schedules, queues, workers, retries and failed-job recovery.
• 🗺️ Cache, search and realtime: caches, search/vector indexes and realtime services when needed.

• 🗺️ Security and abuse controls: access policies, exposed credentials, security headers, rate limits and bot protection. Scoped RLS/advisor and credential-pattern checks exist today.
• 🗺️ Monitoring and alerts: error tracking, logs, uptime and actionable alerts. Provider suggestions are guided today; verified setup is planned. On-demand drift checks exist (golive status, below) — continuous monitoring and alerting do not.
• 🗺️ Product analytics: event validation and consent/data settings, beyond today's guided provider suggestions.
• 🚧 CI/CD and safe releases: previews, release checks, promotion, rollback and drift detection, building on today's approved CLI deployments. Deployment identity — implemented, not live-validated: each successful deploy records the provider's own identity for the deployment it made (deployed:<target>:id = <provider>|<deployment id>|<url>|<time> in .golive/state.json, mock-covered; a provider that reports no identity records none), so a later capability can name one exact deployment. Opt-in preview deploy and release check — implemented, not live-validated: with release.preview: true in golive.yaml (and preview in targets), plan adds preview:deploy — a create that deploys the current working tree to the host's preview target, names the provider, project, env target and the source project the preview shares with production, needs --confirm-live when a live-mode value fills a preview env name, and records the provider's own identity as deployed:preview:id — and release:check, which writes nothing, declares preview:deploy as its prerequisite, and fails the plan when the provider's read of that deployment or a credential scan of its bundle fails. In a cut plan that check is the last step, so what it gates is the promotion (whose own plan re-runs the check before production changes), not a production deploy the plan emits before it. Promotion and rollback — implemented, not live-validated: with release.promote: true (on top of the preview opt-in) a plan asks for a release by promotion, and with release.rollback: true it asks to re-point production at an earlier deployment golive itself created and recorded. promote:production names the exact deployment id it would make production — the provider reports a deployment's id only once the deployment exists, so cutting the candidate and promoting it are two plans and the preview says which one it is — is gated by release:check re-reading that deployment in the same plan, and needs no extra confirmation flag: the plan id, the named deployment and the fresh gate are the approval. Both steps re-read the target deployment and what production serves before writing and prove what production serves afterwards; both keep the cross-release stop (neither is replayable nor a deletion), and neither is automatic — no failed check triggers a rollback, and a deployment built by a dashboard, a Git push or a pull request is never a promotion or rollback target (it is a handoff, named as such). What each host supports differs, and golive refuses rather than guessing: Netlify re-reads its published deployment and can restore an earlier one, so both steps work there; Vercel exposes no read of what production serves and no promote/rollback call golive has exercised, so on Vercel nothing is promoted or rolled back and a warning says why. production-release proves what production serves, names what it served before, and reports a deployment golive never recorded as a handoff. Adding these step ids changes a plan's id, so an approval that was not applied must be re-planned. The preview checks (and therefore promotion) skip on a host that exposes no per-deployment read (Vercel), reporting that instead of guessing. Drift detection exists as the read-only golive status command below (it ran read-only in the auth validation and had nothing actionable once that journey passed, but the DNS, environment, webhook and deployment baselines it compares still lack live evidence), and preview deployments are not yet among the subjects it compares.
• 🗺️ Backups and recovery: retention, restore drills, incident steps and approved cleanup. Approved teardown removes what golive created; backups and any restore remain manual, supervised work.
• ✅ Uninstall / teardown: an approved inventory of golive-created resources and their removal. golive teardown plans the removal, deletes only what golive provably created (ownership proofs and --confirm-destroy), and re-reads the DNS zone's golive-owned record list and the host's own project read after deleting. Nothing it cannot remove is dropped silently: a leftover — an unreadable zone, an unremovable host project, a provider that is not signed in — becomes a handoff naming what remains and the exact fix, and Supabase/Neon projects and the Resend sending domain remain manual handoffs (#9). A removal forgets the baseline golive recorded for that resource, so golive status does not report golive's own teardown as drift, and a sending key golive revoked is reported as a warning instead of a pass, because the provider offers no read to confirm it.
• 🗺️ Costs and quotas: plan choices, budgets, alerts and capacity checks. Scoped Free-plan guards exist today; ongoing cost management is planned.
• 🗺️ Launch essentials: metadata, share previews, indexing, accessibility, support links and owner-reviewed policy pages.
• 🚧 Ownership and handover: accounts, resources, access, renewal responsibilities and maintenance instructions. golive handoff --write records the login route, ownership proofs, recurring jobs and removal gates in GOLIVE_HANDOVER.md, tagging every row as verified, recorded, not verifiable or unknown, and golive status re-reads those subjects on demand: it compares the baselines golive recorded with the providers as they are now, and names what it could not read. Drift re-baselining stays manual and approved — the command is implemented and ran read-only in the auth validation, but a live validation of every drift subject is still pending.

Some steps will always need a person: accepting terms, identity verification, purchases, billing choices and reviews that a provider requires. “Guided” should still mean a clear next action, the right page, the right permissions, a check afterward, and a return to the same workflow. When the app itself needs code changes, GoLive should give the coding agent a concrete task and recheck the result. It should not make you coordinate a dozen disconnected setup conversations.

Next up: complete and live-test the remaining launch journeys—the live run for account isolation (implemented and mock-covered on Supabase, not yet exercised against a real project), live-mode payment flows and the Cloudflare DNS adapter—then expand app architectures and ongoing operations.

These are directions, not release dates. A capability should graduate from experimental only after its account setup, connection, verification and recovery have been exercised. Contributions toward any part of this checklist are welcome, especially evidence of where a real launch gets stuck.

The report records pass, fail, warning and skipped results, along with remaining human steps. Checks include account access, environment-variable names, provider-confirmed deployment URLs, public JavaScript secret patterns, database access, supported auth/webhook/DNS settings, and (where the host can answer it) the deployment the provider says production serves.

A ready deployment is not proof that the app works. An environment-variable name can exist with a wrong value. A verified email domain does not prove inbox delivery. Signed payment events, signup and the app's business flows need functional tests. Skipped is not passed.

Your app gets golive.yaml, .golive/state.json, .golive/report.json and GOLIVE_REPORT.md. State preserves resource IDs and step evidence for recovery; it is not a credential store. golive teardown removes what golive created after its own approval and --confirm-destroy; there is no cross-provider rollback, restore or general reconciliation command for those resources — release:rollback (opt-in) only re-points production at an earlier deployment golive itself recorded, and touches no data, DNS, payment or email resource.

golive handoff --write adds the ownership document: GOLIVE_HANDOVER.md at the repo root and its JSON source in .golive/handover.json. It names the accounts and login route, every resource golive created and the proof it is golive's, what is still manual, what recurs, how removal works, and which commands re-check each subject. Each row says whether it was verified in that run, recorded earlier, not verifiable by golive or unknown; golive read no billing data, so no cost figure is stated. The files contain no secret values, but secret-free metadata can still identify private resources — review them before sharing. --write never overwrites a file golive did not generate unless --force is passed.

golive status asks the follow-up question: has anything changed behind golive's back since it recorded what it did? It compares recorded baselines — the DNS records golive wrote, the environment variable names it delivered, the registered webhook endpoint, the domain attachment, the database project and its connection selectors, the sending domain, the payment account behind the app's keys, the host project — with reads taken now, and labels both sides: expected (recorded by golive <time>) against observed (read now). Each item says who can act: re-run a check, re-plan and apply an approved change, or a decision only a human can make. It is read-only: no report file, no provider write, no state change, and it exits with 2 when something needs acting on. A provider it cannot read is reported as unverifiable — never as clean, and never as a failure — and it never re-baselines anything by itself. Drift is deliberately not a gate: plan, apply and verify never consult it. status is implemented, and it ran read-only during the Supabase auth validation (a failed step surfaced as actionable, then an empty list once it completed), but a live validation of the remaining drift subjects is still pending.

• Approve before account changes. Plans name the destinations and intended writes, and apply refuses without the approved plan id and --yes. Changing the installed release invalidates old approvals. DNS, live-payment and deletion steps have extra gates (--confirm-dns, --confirm-live, --confirm-destroy), and a project's first production deploy needs --confirm-live as well, because approving a plan alone used to be enough to write production for the first time. Trust, access and control walks through each gate.
• Keep secrets out of chat. Supported vendor logins are reused. On macOS, a native hidden-input dialog can save a needed API key; your own editor is the fallback. Keys live in ~/.config/golive/credentials, a local plaintext file at mode 0600 outside your app repo — not an OS keychain, so anything running as your user can read it. The execution code keeps values out of argv, plans, state, reports and command output, storing fingerprints instead of values. golive credentials --remove NAME --yes deletes one stored entry, irreversibly; revoking the token at the provider is what actually ends access. Mac login passwords stay with macOS/vendor authentication prompts; GoLive never asks you to enter one in its key dialog. The full boundary is in trust, access and control.
• Updates have an owner. Skills CLI manages its installs. The optional own installer supports whole-bundle updates and local rollback; automatic replacement is off by default. Update between deployment runs, never between a plan and its apply. Cloud resources are unaffected by rollback.
• Your accounts remain yours. GoLive does not buy services or create billing accounts. Your coding agent, providers and installer have their own data practices.

🤝 We're early, and we'd love your help shaping GoLive. Bug reports, feature ideas, docs fixes and pull requests are all welcome. You don't need to build an adapter to contribute: an unclear login instruction or a real launch that got stuck is useful feedback too.

Open an issue to report a problem or discuss an idea, or send a focused PR. For a larger provider or workflow addition, consider starting an issue so we can agree on the scope together. Never include secrets or raw authentication responses in a report.

See CONTRIBUTING.md for local setup, tests and your first contribution. The architecture, trust, access and control, recovery, provider scope and validation record explain what exists and where help is needed.

MIT licensed. Bundled third-party notices are included in THIRD_PARTY_NOTICES.md.