Vercel deployment

The demo runs on Vercel’s free tier. The Vercel CLI takes care of build, deploy, and domain configuration.

Steps

1. Fork or clone the repo

gh repo fork bgood11/lending-agent --clone
cd lending-agent

Or clone directly if you don’t need a fork:

git clone https://github.com/bgood11/lending-agent
cd lending-agent

2. Install the Vercel CLI

pnpm install -g vercel

3. Link to a Vercel project

vercel link

You’ll be asked which scope (personal or team) and which project (or to create a new one). Choose your preferred scope and let it create a new project named lending-agent (or whatever you prefer).

This creates a .vercel/project.json file linking your local checkout to the Vercel project. It’s gitignored.

4. Set the API key as a production environment variable

vercel env add ANTHROPIC_API_KEY production

Paste your Anthropic API key when prompted. The CLI will encrypt and store it server-side.

5. Deploy

vercel deploy --prod

The first deploy takes 1-2 minutes (build + edge propagation). You’ll get a URL like https://lending-agent-yourname.vercel.app.

Open it. Walk a journey to confirm.

Auto-deploys on push

Vercel can auto-deploy on every push to your fork’s main branch. To enable:

vercel git connect

This connects the Vercel project to your GitHub fork. After connecting, every push to main triggers a production deploy. Pushes to other branches create preview deploys at https://lending-agent-git-<branch>-yourname.vercel.app.

Cost expectations

Per customer journey:

~10-20 model calls. Each turn is roughly 2-3K input tokens (system prompt + conversation history + system notes) and up to 900 output tokens.
At current Claude Sonnet 4.6 list rates (input $3/MTok, output $15/MTok), a single turn is ~£0.01. A full journey lands around £0.10-£0.20 in raw model spend.

The replay engine adds calls on demand. A 5-run replay across 3 disclosures is 15 model calls at ~£0.005 each, so roughly £0.07 per replay run.

For a demo deployment with low traffic, expect monthly costs in single-digit pounds. For a production deployment processing hundreds of journeys a day, model calls are the dominant cost line. Prompt caching on the system prompt (shared across every turn) cuts the input bill substantially; see Production hardening for the longer cost-control list.

These figures are list prices and will move. Always check current Anthropic pricing before quoting a cost figure to a stakeholder.

Vercel’s “personal” vs “team” scopes

Vercel has two scope types: personal (free hobby tier) and team. The demo runs fine on personal.

If you want commercial use (custom domains beyond personal-account limits, enhanced support, more aggressive caching), move the project to a team. The CLI flow is:

# Switch the active scope to your team (one-off, persisted in ~/.vercel)
vercel switch your-team-slug

# Re-link the project under the team scope
vercel link --scope your-team-slug

If the team doesn’t exist yet, create it via the dashboard at vercel.com/teams; the CLI doesn’t create teams.

Build configuration

Out of the box, Vercel detects Next.js and configures the build automatically. No vercel.json needed.

If you want to customise:

buildCommand: defaults to pnpm build
installCommand: defaults to pnpm install
outputDirectory: defaults to .next
framework: detected as nextjs

If you need to customise these, use vercel.json at the repo root or set them in the project settings on the Vercel dashboard. The demo doesn’t ship a vercel.json because the auto-detected defaults are correct.

Logs and monitoring

The Vercel dashboard surfaces:

Function logs: every server-side request, with status codes and durations
Runtime logs: console output from server functions
Build logs: per-deploy build output
Analytics: traffic, top pages, top function invocations (paid tier)

For debugging the agent specifically: open a customer URL with ?demo=1, walk a journey, and check Function logs for /api/chat/customer invocations. The chat route logs Anthropic call failures with stack traces.

Custom domain

Once you’ve got a working deploy, add a custom domain:

vercel domains add yourdomain.co.uk

The CLI prints the DNS records you need to add (typically a CNAME for www and an A record or ALIAS for the apex). Follow the instructions for your DNS provider; most updates propagate in a few minutes.

See Custom domain for the full DNS configuration.

CI

Vercel’s auto-deploy is enough CI for a demo. For production:

Linting and type-checking on PR: GitHub Actions running pnpm lint && pnpm build before merge.
Replay regression: a corpus of historical cases stored in CI, replayed against any model or prompt change. See Audit & replay integration.
Smoke tests on staging: deploy to a preview, run bash scripts/smoke.sh BASE=$PREVIEW_URL, gate prod promotion on the result.

Rolling back

If a deploy breaks something:

vercel rollback

Vercel maintains the previous production deploy as a hot standby. Rollback is instant. The bad deploy is still archived; you can inspect it via the dashboard.

For more controlled rollouts (canary deploys, gradual rollout), Vercel’s Rolling Releases feature handles this; the demo is too small to need it.