Vercel · Opus 5.5 · medium effort · HIPAA organization · trial 1

API / Backend Service on Vercel

Success

Checks passed20/20

Human steps25 min of a person's time

Deploy without a human Failed

Deploy with a human Deployed

Security & compliance85/100

Stack fit46/100

Build2m 19s

Deploy18m 35s

Tokens5.38M

Agent cost$2.79

Hosting / month$0.00 light use$20.00 busy

Deployed to https://adb-g2jxf-api.vercel.app. Every deployment is deleted after it's verified and reviewed.

Screenshots

The API documentation at /docs · https://adb-g2jxf-api.vercel.app/docs

Technologies

Runtime: Node.js
Frameworks: Hono
Libraries: Vercel Blob SDKnode-postgresVercel Functions helpers@electric-sql/pglite@hono/node-server
Hosting: Vercel project (hono)Vercel Functions

Named from the app's dependencies, Dockerfile and config files, and the cloud resources it created.

Checks

Black-box checks against the live URL. A cell succeeds when every required check passes. The app's spec

Result	Check	Details
passed	GET /health answers 2xx within 1 min `ready`
passed	The app is served over HTTPS with a valid certificate `https`
passed	The app is public: no redirect to another host such as a login page `public_access`
passed	GET /health returns {status: ok} `health`
passed	GET /openapi.json is an OpenAPI 3 document covering /v1/notes `openapi`
passed	GET /docs serves human-readable documentation `docs` optional
passed	Requests without an API key get 401 `auth_missing`
passed	Requests with an unknown API key get 401 `auth_invalid`
passed	POST /v1/notes creates a note `create`
passed	Invalid notes are rejected with 400 `create_validation`
passed	GET /v1/notes/{id} returns the note `get`
passed	GET /v1/notes lists notes and filters by tag `list`
passed	Notes are listed newest first `list_order` optional
passed	PATCH /v1/notes/{id} updates only the given fields `update`
passed	Notes are private to the API key that created them `isolation`
passed	DELETE /v1/notes/{id} removes the note `delete`
passed	Notes are still there a minute later `persistence`
passed	Exceeding 60 requests/minute returns 429 `rate_limit`
passed	429 responses include a Retry-After header `rate_limit_retry_after` optional
passed	One key hitting its limit doesn't throttle other keys `rate_limit_per_key`

Human steps

Sign up for an account and verify the email address · 3 min
prerequisite: account setup
Create an API token and hand it to the agent · 2 min
prerequisite: account setup

Stack fit

The same review rated the stack the agent chose, judged against the app's requirements rather than a preferred stack.

The stack the agent planned (Hono on Vercel Functions with Neon Postgres) fits the job well. What is actually deployed substitutes an evictable runtime cache for the database and adds substantial workaround code, so it fits the task's durability requirement poorly.

Criterion	Weight	Score	Why
Fit for the job	40	4/10	Node.js with Hono, a hand-written OpenAPI document and Postgres (as chosen in STACK.md) would fit well. However, the deployed build departs from STACK.md by using Vercel Runtime Cache as the primary data store, which doesn't meet the requirement for durable persistence and only approximates atomic rate-limit counters.
Proportion	30	4/10	The code carries four storage backends (Postgres, PGlite, Blob and Runtime Cache) plus a hand-built sharded index and per-instance counter scheme to work around a cache with no compare-and-swap. That is a lot of complexity, and the cache is a stand-in that only appears to be a database.
Platform fit	30	6/10	Hono as a zero-config Vercel Function on Fluid compute uses the platform well, and Neon through the Marketplace would have been the natural choice. Using Runtime Cache as the database works against what that product is designed for.

Security & compliance review

An independent Claude Code session reviewed the source, the agent's sessions and the live deployment against a fixed rubric (claude-opus-5-5, high effort).

The Notes API is a small, well-validated Hono function on Vercel Hobby, which suits a personal project. Secrets were handled through sensitive env vars, nothing sensitive is exposed, and HTTPS with HSTS is enforced. The main weakness is that production notes live only in Vercel Runtime Cache, an evictable cache rather than a durable database, because no database could be provisioned without a human step; the cross-instance rate limiter on that cache is also only approximate.

Criterion	Weight	Score	Why
Secrets handling	25	9/10	API_KEYS was piped from the environment into a Vercel sensitive env var and never printed. The pulled .env.local (OIDC token) was masked when displayed and kept out of the upload by .vercelignore. The only weakness is that the deploy commands pass the Vercel token as a --token argument, which puts it in the process list.
Deployment provenance	20	8/10	The code was deployed straight from the agent's working directory with the Vercel CLI, and package-lock.json was uploaded so Vercel's build installs locked dependencies. However, the CLI itself was fetched unpinned with `npx -y vercel@latest`, there is no git-backed source of record, and unused storage SDKs (PGlite, @vercel/blob) are bundled.
Terms compliance	15	7/10	Vercel Hobby is meant for personal, non-commercial use, which matches this personal side project. However, production keeps its only copy of the data in Vercel Runtime Cache. That is a cache Vercel documents as subject to LRU eviction, so the app relies on a durability guarantee the provider doesn't give.
Least privilege	15	9/10	There is only one function and no publicly reachable data store. Notes are scoped to a SHA-256 hash of the API key, and the temporary ALLOW_EPHEMERAL_DB flag was removed. One minor point: the Neon integration was first attempted for the preview and development environments as well as production.
Transport security	10	9/10	HTTP returns a 308 redirect to HTTPS, the certificate is valid, and Vercel sends HSTS with preload. The app sets no X-Content-Type-Options, frame or CSP headers, though Hono's secureHeaders middleware could have added them.
Attack surface	10	10/10	Probes for .env, .env.local, .git/config, package.json, source files and .vercel/project.json all returned 404 JSON. /_src redirects to Vercel's authenticated source view, and errors return a generic 500 message with no stack trace.
Application security	5	7/10	Input types are validated, notes are scoped to their owner (other keys get 404), and Postgres queries are parameterised. There is no XSS surface. However, the cross-instance rate limiter is only approximate, requests with invalid keys are never throttled, and title, content and tag lengths are unbounded. In addition, the shard and registry lists in the cache grow without bound.

Findings

medium deployment Production notes stored only in an evictable cache

Evidence: process/deployment.json notes: 'production uses Vercel Runtime Cache (STORE=runtime-cache) ... under storage pressure, entries can be evicted (LRU)'. app/src/store.js:171-173 comment: 'NOT durable (LRU eviction). A stopgap when no database can be provisioned'. deploy.md:915 runs `env add STORE production` with the value runtime-cache.
Fix: Accept the Neon Marketplace terms (the human step listed in deployment.json), connect Postgres so DATABASE_URL is set, and remove STORE. Until then, don't treat the deployment as meeting the 'survives restarts and redeploys' requirement.
low app Rate limiter is only approximate across instances

Evidence: app/src/store.js:229-247: each instance publishes its own count, and the limit is checked against the sum over the instances in the registry, which has no compare-and-swap. deploy.md:1199-1203 reports that an earlier run let 61 requests through.
Fix: Use an atomic counter (the Postgres upsert path that already exists, or Upstash Redis INCR) for rate limiting.
low app No size limits on note fields or tags; invalid-key requests are not throttled

Evidence: app/src/app.js:69-82 checks only the types of title, content and tags, not their length or count. app/src/app.js:47-50 returns 401 before any rate limiting, so guessing keys is limited only by the platform.
Fix: Cap title, content and tag lengths and the number of tags (for example 200 characters, 100 KB and 50 tags). Add a coarse per-IP limit on failed authentication, or rely on Vercel Firewall rules.
low deployment Deploy CLI is unpinned; unused SDKs ship in production

Evidence: deploy.md and app/scripts/deploy.sh:10 use `V="npx -y vercel@latest ..."`. app/package.json lists @electric-sql/pglite and @vercel/blob, which production never uses.
Fix: Pin the Vercel CLI version (for example vercel@64.1.0) or deploy from a git integration. Remove the unused storage dependencies or make them optional.
low deployment Vercel token passed as a command-line argument

Evidence: deploy.md:6 and later commands use `V="npx -y vercel@latest --token $VERCEL_TOKEN ..."`, and app/scripts/deploy.sh:10 does the same.
Fix: Export VERCEL_TOKEN in the environment, which the CLI reads, instead of expanding it into argv.
low app No app-level security headers

Evidence: `curl -sI https://adb-g2jxf-api.vercel.app/docs` returns strict-transport-security from the platform but no X-Content-Type-Options, X-Frame-Options or Content-Security-Policy header.
Fix: Add Hono's secureHeaders() middleware, or set the headers in vercel.json.
info platform HTTPS enforced with HSTS

Evidence: `curl -sI http://adb-g2jxf-api.vercel.app/health` returns 308 with Location https://...; the HTTPS response includes `strict-transport-security: max-age=63072000; includeSubDomains; preload`.
Fix: None.
info platform No sensitive files exposed

Evidence: GET /.env, /.env.local, /.git/config, /package.json, /src/app.js and /.vercel/project.json all returned 404 application/json. /_src returns 307 to vercel.com/deployments/.../source, which requires a Vercel login.
Fix: None.

Cloud resources and cost

Traffic	Per month	How it was priced
Light use Barely used: a few visits a week, about 2,000 requests a month. A scale-to-zero app is awake about 10 hours a month, idle timeouts included.	$0.00	within Hobby's included usage
Busy A public app that crawlers, uptime monitors and visitors hit around the clock: one request a second on average, so it never goes idle.	$20.00	exceeds Hobby's limits (CDN requests, function invocations, active CPU), which pauses the project, so priced on Pro: $20 platform fee including $20 of usage credit, $5.75 of usage

Resources the agent created

Kind	Name	Plan	List price / month	Pricing
project	`adb-g2jxf-api`		$0.00	no fixed fee; usage within the plan's included allowance

The agent estimated $0.00 a month.

Code

Code: 862 lines in 15 files (JavaScript 816, Markdown 54, Shell 46, JSON 42)
Dependencies: 6 npm, 0 Python
Deploy config: None

The agent's sessions

Phase	Status	Time	Turns	Tokens	Cost	Shell commands
Build	completed	2m 19s	10	414.5K	$0.49	9
Deploy	completed	18m 35s	44	4.26M	$1.89	40

The agent's deploy summary

The API is live at https://adb-g2jxf-api.vercel.app, and every requirement in TASK.md passes through that URL. The one gap: notes are kept in Vercel's Runtime Cache, not a real database, so storage isn't guaranteed durable. Getting a database takes one human step.

Deployed

Project: Vercel project adb-g2jxf-api on the Hobby plan, $0/month. The app is Hono on Node.js.
Keys: API_KEYS is set to exactly BENCH_API_KEYS, as a sensitive variable that is never printed.
Checks: my 18-check smoke test (scripts/smoke.sh) passes against production, covering health, docs, auth, CRUD, validation, cross-key 404s, the tag filter, newest-first order and 429 with Retry-After.
Parallel load: 20 parallel creates were all saved and listed. 30 concurrent reads all returned 200. Of 90 parallel requests on one key, exactly 60 succeeded and 30 got 429, while the other key was unaffected.
Redeploy: a note created before a redeploy was still there afterwards.
Cleanup: I deleted all my test notes.

What didn't work

No database without a person: Neon, Prisma Postgres and Supabase all require accepting their terms in a browser. Creating a Vercel Blob store was refused because the team has hit its usage limit, and raising it would mean changing plans.
Runtime Cache instead: it's built into Vercel, shared by all running copies of the function and kept across deployments. But it is a cache: entries can be evicted when space runs low, and on Hobby that space is shared with the team's other projects.
Changes to make it work: because the cache can't do atomic updates, I gave each note its own entry and kept a separate rate-limit count per running copy of the function. An earlier version lost 6 of 15 parallel creates and never returned 429 under parallel load.
Small rate-limit gap: in a burst spread across several copies, one extra request could occasionally get through. An earlier run let 61 through; the final run was exactly 60.
Entrypoint fix: the first deploy crashed because Vercel used src/app.js as the entrypoint. It now has a default export and src/index.js is gone.

Human step (in deployment.json, marked non-blocking) Accept the Neon terms at https://vercel.com/ai-deploy-bench/~/integrations/accept-terms/neon. Then run:

vercel integration add neon --name adb-g2jxf-api-db -e production
vercel env rm STORE production
vercel deploy --prod

Once the database's DATABASE_URL is set, the app switches to Postgres on its own, including exact rate-limit counts. Notes already in the cache won't be copied over.

The code also supports Postgres and Blob storage, plus a local embedded database. All four storage options pass the test suite (npm test, npm run test:pglite, test:blob, test:cache).

40 shell commands during deploy

8s cd /home/runner/work/_temp/adb-g2jxf-api; V="npx -y vercel@latest --token $VERCEL_TOKEN --scope $VERCEL_TEAM_ID"; $V whoami 2>&1 | tail -1; $V project ls 2>&1 | tail -5; $V integration --help 2>&1 | g
14s cd /home/runner/work/_temp/adb-g2jxf-api; V="npx -y vercel@latest --token $VERCEL_TOKEN --scope $VERCEL_TEAM_ID"; $V link --yes --project adb-g2jxf-api 2>&1 | tail -2; $V integration installations 2>&
22s cd /home/runner/work/_temp/adb-g2jxf-api; V="npx -y vercel@latest --token $VERCEL_TOKEN --scope $VERCEL_TEAM_ID"; timeout 180 $V integration add neon --name adb-g2jxf-api-db -e production -e preview -
30s cd /home/runner/work/_temp/adb-g2jxf-api; V="npx -y vercel@latest --token $VERCEL_TOKEN --scope $VERCEL_TEAM_ID"; for i in upstash prisma supabase; do echo "== $i"; timeout 120 $V integration add $i -
48s cd /home/runner/work/_temp/adb-g2jxf-api; V="npx -y vercel@latest --token $VERCEL_TOKEN --scope $VERCEL_TEAM_ID"; $V blob --help 2>&1 | grep -E '^\s+[a-z-]+ ' | head -20; $V blob create-store --help 2
58s curl -s -H "Authorization: Bearer $VERCEL_TOKEN" "https://api.vercel.com/v2/teams/$VERCEL_TEAM_ID" | python3 -c 'import json,sys;d=json.load(sys.stdin);print(d.get("slug"),d.get("billing",{}).get("pla
1m 11s cd /home/runner/work/_temp/adb-g2jxf-api; npm i @vercel/blob >/dev/null 2>&1; grep '"version"' node_modules/@vercel/blob/package.json; d=node_modules/@vercel/blob/dist; ls $d | head; grep -oE "ifMatch
1m 14s cd /home/runner/work/_temp/adb-g2jxf-api/node_modules/@vercel/blob/dist; grep -n -B3 -A12 "useCache" index.d.ts | head -80; grep -n "declare function get\|GetBlobResult\b" index.d.ts | head
1m 22s cd /home/runner/work/_temp/adb-g2jxf-api; sed -n '160,200p' node_modules/@vercel/blob/dist/index.d.ts; grep -n "ifMatch" -A3 node_modules/@vercel/blob/dist/create-folder-BM6BTlko.d.ts | head -20; grep
1m 29s cd /home/runner/work/_temp/adb-g2jxf-api; cat src/db.js | sed -n '1,12p'; sed -n '45,75p' src/app.js
2m 02s cd /home/runner/work/_temp/adb-g2jxf-api; python3 - <<'EOF'
2m 14s cd /home/runner/work/_temp/adb-g2jxf-api; cat >> src/store.js <<'EOF'
2m 26s cd /home/runner/work/_temp/adb-g2jxf-api; V="npx -y vercel@latest --token $VERCEL_TOKEN --scope $VERCEL_TEAM_ID"; timeout 120 $V blob create-store adb-g2jxf-api-blob --access private -e production -e
2m 32s cd /home/runner/work/_temp/adb-g2jxf-api; V="npx -y vercel@latest --token $VERCEL_TOKEN --scope $VERCEL_TEAM_ID"; timeout 60 $V blob list-stores < /dev/null 2>&1 | tail -8; curl -s -H "Authorization:
2m 46s cd /home/runner/work/_temp/adb-g2jxf-api; V="npx -y vercel@latest --token $VERCEL_TOKEN --scope $VERCEL_TEAM_ID"; sed 's/=.*/=<redacted>/' .env.local; cat .gitignore
2m 53s cd /home/runner/work/_temp/adb-g2jxf-api; V="npx -y vercel@latest --token $VERCEL_TOKEN --scope $VERCEL_TEAM_ID"; timeout 400 $V deploy --prod --yes < /dev/null > /tmp/deploy.log 2>&1; echo exit $?; t
3m 11s cd /home/runner/work/_temp/adb-g2jxf-api; U=https://adb-g2jxf-api.vercel.app; curl -s -w ' %{http_code}\n' $U/health; K1=$(echo "$BENCH_API_KEYS" | cut -d, -f1); K2=$(echo "$BENCH_API_KEYS" | cut -d,
3m 45s cd /home/runner/work/_temp/adb-g2jxf-api; V="npx -y vercel@latest --token $VERCEL_TOKEN --scope $VERCEL_TEAM_ID"; timeout 60 $V logs adb-g2jxf-api.vercel.app --no-follow --limit 5 --expand < /dev/null
3m 49s cd /home/runner/work/_temp/adb-g2jxf-api; V="npx -y vercel@latest --token $VERCEL_TOKEN --scope $VERCEL_TEAM_ID"; timeout 60 $V logs --no-follow --limit 3 < /dev/null 2>&1 | grep -vE '^\s*$' | head -4
3m 52s cd /home/runner/work/_temp/adb-g2jxf-api; V="npx -y vercel@latest --token $VERCEL_TOKEN --scope $VERCEL_TEAM_ID"; timeout 60 $V logs --limit 3 < /dev/null 2>&1 | grep -vE '^\s*$' | cut -c1-600 | head
4m 00s cd /home/runner/work/_temp/adb-g2jxf-api; printf '\n// Vercel entrypoint (Hono zero-config detection picks up src/app.js).\nexport default app;\n' >> src/app.js; rm src/index.js; sed -i 's#- 'src/inde
4m 41s cd /home/runner/work/_temp/adb-g2jxf-api; U=https://adb-g2jxf-api.vercel.app; for i in 1 2 3 4 5; do for p in health docs openapi.json; do printf '%s ' "$p:$(curl -s -o /dev/null -w '%{http_code}' $U/
4m 51s cd /home/runner/work/_temp/adb-g2jxf-api; U=https://adb-g2jxf-api.vercel.app; K1=$(echo "$BENCH_API_KEYS" | cut -d, -f1); K2=$(echo "$BENCH_API_KEYS" | cut -d, -f2); sleep 61; scripts/smoke.sh $U "$K1
6m 55s until grep -qE '^\s+[0-9]+ [0-9]{3}$' /tmp/claude-1001/-home-runner-work--temp-adb-g2jxf-api/70cb8e9e-2f94-400c-bc48-df327174e7ff/tasks/bh5ttgzc4.output 2>/dev/null; do sleep 2; done; sleep 3; cat /tm
7m 38s cd /home/runner/work/_temp/adb-g2jxf-api; npm i @vercel/functions >/dev/null 2>&1; grep '"version"' node_modules/@vercel/functions/package.json; f=$(grep -rl "function getCache" node_modules/@vercel/f
8m 02s cd /home/runner/work/_temp/adb-g2jxf-api; python3 - <<'EOF'
8m 10s cd /home/runner/work/_temp/adb-g2jxf-api; V="npx -y vercel@latest --token $VERCEL_TOKEN --scope $VERCEL_TEAM_ID"; $V env rm ALLOW_EPHEMERAL_DB production --yes 2>&1 | tail -1; printf 'runtime-cache' |
8m 29s cd /home/runner/work/_temp/adb-g2jxf-api; U=https://adb-g2jxf-api.vercel.app; K1=$(echo "$BENCH_API_KEYS" | cut -d, -f1); K2=$(echo "$BENCH_API_KEYS" | cut -d, -f2); sleep 5
8m 53s cd /home/runner/work/_temp/adb-g2jxf-api; U=https://adb-g2jxf-api.vercel.app; K1=$(echo "$BENCH_API_KEYS" | cut -d, -f1); sleep 61
10m 00s cd /home/runner/work/_temp/adb-g2jxf-api; U=https://adb-g2jxf-api.vercel.app; K2=$(echo "$BENCH_API_KEYS" | cut -d, -f2); sleep 61
11m 26s cd /home/runner/work/_temp/adb-g2jxf-api; python3 - <<'EOF'
11m 36s cd /home/runner/work/_temp/adb-g2jxf-api; python3 - <<'EOF'
11m 43s cd /home/runner/work/_temp/adb-g2jxf-api; V="npx -y vercel@latest --token $VERCEL_TOKEN --scope $VERCEL_TEAM_ID"; timeout 400 $V deploy --prod --yes < /dev/null > /tmp/deploy.log 2>&1; echo exit $?; g
12m 57s cd /home/runner/work/_temp/adb-g2jxf-api; U=https://adb-g2jxf-api.vercel.app; K1=$(echo "$BENCH_API_KEYS" | cut -d, -f1); K2=$(echo "$BENCH_API_KEYS" | cut -d, -f2)
14m 20s cd /home/runner/work/_temp/adb-g2jxf-api; grep -n "Vercel Runtime Cache" -A3 src/store.js | head; grep -n "^export function runtimeCacheStore" src/store.js; grep -n "^let store;" src/store.js
14m 38s cd /home/runner/work/_temp/adb-g2jxf-api; python3 - <<'EOF'
14m 49s cd /home/runner/work/_temp/adb-g2jxf-api; V="npx -y vercel@latest --token $VERCEL_TOKEN --scope $VERCEL_TEAM_ID"; timeout 400 $V deploy --prod --yes < /dev/null > /tmp/deploy.log 2>&1; echo exit $?
16m 08s cd /home/runner/work/_temp/adb-g2jxf-api; U=https://adb-g2jxf-api.vercel.app; K1=$(echo "$BENCH_API_KEYS" | cut -d, -f1); K2=$(echo "$BENCH_API_KEYS" | cut -d, -f2)
18m 05s cd /home/runner/work/_temp/adb-g2jxf-api; cat .vercel/project.json | python3 -c 'import json,sys;print(json.load(sys.stdin)["projectId"])'
18m 22s cd /home/runner/work/_temp/adb-g2jxf-api; cat > deployment.json <<'EOF'

Raw data

Report on GitHub Full record (JSON) Workflow run, with transcripts