ScopeCard
ScopeCard scans a repo's AI agent tools and MCP configs and publishes a hash-stamped capability card plus a pre-filled answer pack, so a small SaaS can clear an agent security review in an afternoon.
One product, one founder, thirty days. Everything below is built to be executed inside the contest window — the official rules still decide what counts.
Decide in 20 seconds
The facts that decide whether this is worth thirty days of yours.
Technical founder or lead engineer at a 1-20 person B2B SaaS that has shipped an AI agent or MCP-backed feature and is blocked on a prospect's security questionnaire.
Market demand 4.0/10
Narrow the buyer to a segment with an urgent, already-budgeted problem and reference the specific trigger event that creates willingness to pay.
Why this could work in 30 days: one founder can ship the MVP in 5 days; the plan targets a first sale on day 12; 11 sales at $99 clears $1,000 of collected revenue; strongest signal is speed to first revenue at 7.7/10.
Score and projections come from the generated report and are not guarantees. The $100K Hackathon is run by Ship or Die — read the official rules before you build.
Your first 24 hours
20 concrete actions. Tick them off as you go — nothing is saved or sent.
Product opportunity and positioning
Target customer
Technical founder or lead engineer at a 1-20 person B2B SaaS that has shipped an AI agent or MCP-backed feature and is blocked on a prospect's security questionnaire.
Why now
Small SaaS are shipping agents faster than their buyers' security teams can review them, so agent capability questionnaires are arriving now and every one stalls a deal.
Problem
Agent data-access questions arrive as a 20-40 question security questionnaire; the founder re-reads tool schemas and MCP configs, writes a fresh document per prospect, and still cannot prove the answer matches the deployed commit.
Solution
Connect a GitHub repo or upload a zip; ScopeCard classifies every tool as read/write/destructive/external-network, publishes a hosted card stamped with commit SHA and content hash, and exports a JSON manifest, scopecard.yaml lockfile and PDF/DOCX answer pack.
1. Product opportunity and positioning
An AI agent is a set of tools with permissions. Buyers of small B2B software have started asking for proof of that surface inside security reviews: which data can the agent reach, which actions can it take, which credentials does it hold, where does a human approve. The founder answering today re-reads tool schemas and mcp.json by hand, writes fresh prose per prospect, and cannot show the answer matches the deployed commit.
ScopeCard converts the repo into the answer. Connect a GitHub repo or upload a zip; ScopeCard parses MCP config files and tool/function schemas, classifies each tool as read, write, destructive or external-network with a confidence flag, and publishes a hosted capability card stamped with the scanned commit SHA and a content hash. The same scan emits a JSON manifest, a scopecard.yaml lockfile and a PDF/DOCX answer pack mapped to the roughly fifteen questions agent reviews actually ask.
Positioning is deliberately narrow: not a runtime proxy, not observability, not an identity fabric. The buyer is the technical founder at a 1-20 person B2B SaaS with one shipped agent and one prospect blocking the deal. The trigger is a dated questionnaire, the approval is a single card payment, and the deliverable is a document generated from code in one session, so value lands before the invoice clears.
JEV scores speed to first revenue highest (7.72 normalized) and market demand lowest (3.98). That is the whole strategy in one line: the product is fast to sell if the buyer exists, so treat demand as a hypothesis to be tested by day 12 rather than an assumption to be celebrated in a pitch deck. The weakest scored axis becomes the first experiment, not the last section of the report.
2. Smallest sellable MVP
Five things must exist before the first dollar. Repo ingest via GitHub OAuth or zip upload. A classifier that labels every tool read, write, destructive or external-network with confidence, plus a UI override. A hosted card at a stable slug showing commit SHA and content hash. A paid unlock that removes the watermark and releases the manifest and PDF/DOCX answer pack. Stripe Checkout. Everything else is an upgrade shipped after day 12: lockfile CLI diff, re-scan webhooks, custom logo, multi-project.
The test of MVP completeness is single: can a stranger open a preview URL, recognise their own tool list, pay $99, and forward the resulting PDF to their prospect within ten minutes without talking to the founder. Nothing about runtime interception ships, ever, in this window.
3. Technical implementation plan
Next.js on Vercel, Supabase Postgres, Octokit, Anthropic API, react-pdf, Stripe Checkout, Resend. Ingest is a serverless route that pulls the tree, matches known config filenames plus any JSON or YAML file containing tool or function schemas, and normalises to a tool inventory keyed by name. Classification is one batched Anthropic call per scan with strict JSON output and one retry, cached by commit SHA so re-renders cost nothing.
The card page renders from the stored inventory plus the hash; the manifest and lockfile are serialisations of the same record, so they cannot disagree. Payment flips a published flag scoped to one project for 90 days. The hardest technical risk is parser recall on monorepos and JSONC. Mitigate by shipping an explicit "tools we could not classify" bucket on the card instead of silently dropping entries.
4. Pricing and monetization
Three one-time tiers: Card $99, Review-Ready $249, Studio $499. One-time is correct because the job is a single review, but the repeat path is real: every new prospect questionnaire and every deploy that changes the tool surface creates a re-publish or upgrade event, and Studio turns one agency relationship into five projects.
Processing costs roughly 4% (2.9% plus $0.30 plus platform), classification costs under $0.25 per card, so margin stays above 90% at $99. Do not discount below $99. Instead give the first three buyers a public worked example in exchange for permission to describe the questionnaire they answered.
5. First-customer acquisition
One channel: GitHub code search for "mcpServers" plus tool-schema files, filtered to repos pushed in the last 90 days, then one personal email per maintainer containing a watermarked preview card for their own public repo. The email leads with their numbers, not a pitch.
Target 360 sends across seven waves with one 72-hour follow-up each, and a hard read at day 21: fewer than three paid publishes from 240 sends means the channel or the offer is wrong, and the plan switches to warm repliers and the agency angle rather than adding volume to a dead list.
6. Thirty-day execution roadmap
Days 1-5 build to a live checkout. Day 7 verifies a real purchase and the TrustMRR connection. Days 8-12 generate 200 preview cards and send 180 emails, with day 12 as the first-sale target. Days 13-18 convert replies, ship the lockfile and re-scan upgrades, publish one anonymised worked example, and post in developer communities asking specifically about classification accuracy. Days 19-21 run a robustness pass and the channel review. Days 22-27 add an agency batch, a methodology page to cut refund risk, and two more waves. Days 28-30 reconcile Stripe against TrustMRR, run the final push, and write the post-mortem. Every day has one deliverable and no day carries more than four tasks.
7. Revenue and profit scenarios
Projections only, derived from stated assumptions, not verified revenue. Conservative: 3 sales at $99 is $297 gross, $12 processing, $25 operating, $285 hackathon profit, $260 business profit. Base: 10 sales is $990 gross, $39 processing, $45 operating, $951 hackathon profit, $906 business profit. Upside: 25 sales is $2,475 gross, $50 refunds, $97 processing, $150 sponsored community slots, $90 operating, $2,178 hackathon profit, $2,088 business profit.
Base implies roughly 10 paid publishes from 360 personalised sends, about 2.8% conversion, which has not been observed. The contest ranks hackathon profit, which ignores hosting and API delivery cost; the business column is what the founder actually keeps.
8. Main risks and validation experiments
Parser recall: if cards miss dynamically registered tools the artifact is worse than useless. Test by scanning 20 real public MCP repos by day 6 and counting unclassified entries; widen the parser or narrow the promise if more than 30% of tools are unclassified on half the sample.
Demand: cold GitHub reply rates may sit under 3%. Test with 60 sends by day 10 and measure opens and replies by day 12; if replies are under three from 180 sends, re-cut the offer around the questionnaire deadline or move to agencies.
Willingness to pay for a document rather than software: watch for refund requests citing "expected enforcement". Put the limitation on the card page and pricing FAQ by day 26 and track refund rate; above 10% of sales, the positioning is wrong. Two secondary risks: mislabelling a destructive tool as safe, mitigated by confidence flags, manual override and never calling a card verified safe; and Anthropic rate limits during an outreach push, mitigated by backoff and per-commit caching.
9. First 24-hour action plan
Hours 1-3: repo, Supabase schema, Stripe test prices, Vercel skeleton deployed. Hours 3-6: GitHub OAuth and file detection working against two known public repos. Hours 6-10: classification prompt producing validated JSON for those repos, inventory stored per commit. Hours 10-14: card page rendering commit SHA, hash and tool table. Hours 14-18: Stripe Checkout, watermark gate and PDF export, deployed. Hours 18-24: run the flow end to end on one unfamiliar repo, fix the two worst failures, and generate the first three preview cards for real repos to seed day-10 outreach.
Smallest sellable MVP
Repo connect, tool classification with override, a hosted hash-stamped card, Stripe-paid unlock that releases the manifest and PDF answer pack.
Must-have features
- GitHub OAuth or zip upload with detection of mcp.json, claude_desktop_config.json and .cursor/mcp.json
- Tool/function schema parsing into a normalised inventory with an unclassified bucket
- LLM classifier producing read/write/destructive/external-network labels with confidence and manual override
- Hosted capability card at a stable slug showing scanned commit SHA and content hash
- JSON manifest download and PDF/DOCX answer pack mapped to ~15 standard agent security questions
- Stripe Checkout with a watermarked free preview and paid publish unlock
- Resend delivery email containing the card URL, manifest and answer pack
Explicitly out of scope
- Runtime proxy or interception of live agent calls
- Slack or email approval buttons
- Traffic logging, SIEM or observability
- SOC 2, SSO, RBAC or admin console
- Self-hosted deployment
- Agent orchestration or agent building
- Multi-project billing before day 24
Demo flow (what the buyer sees)
- Buyer opens a preview card URL generated from their public repo
- Card shows tool count, destructive count, commit SHA and content hash
- Buyer clicks Publish and pays $99 via Stripe Checkout
- Watermark is removed and the project is marked published for 90 days
- Buyer downloads manifest JSON and the PDF/DOCX answer pack
- Buyer forwards the card URL and PDF to the prospect's security reviewer
Technical implementation plan
Stack
Architecture
A serverless ingest route pulls the repo tree, matches known MCP config filenames plus any JSON/YAML file containing tool or function schemas, and writes a normalised tool inventory keyed by name and commit SHA into Supabase. One batched Anthropic call per scan classifies each tool with strict JSON output, cached by commit hash; the founder can override any label in the UI. The card page, JSON manifest and scopecard.yaml lockfile are all rendered from that one stored record plus its content hash, so they cannot disagree. Stripe Checkout sets a published flag scoped to one project for 90 days, and Resend emails the card URL, manifest and PDF.
Build steps
- Scaffold Next.js with Supabase schema for projects, scans, tools, cards and purchases
- Add GitHub OAuth read access and a zip-upload fallback
- Implement config-file detection and a normalised tool inventory with an unclassified bucket
- Write the Anthropic classification prompt with JSON schema validation and one retry
- Build the hosted card page with commit SHA, content hash and tool table
- Add manual label override and re-render of the card
- Serialise the same record into manifest JSON and scopecard.yaml
- Generate the PDF/DOCX answer pack with react-pdf from a 15-question template
- Wire Stripe Checkout and the watermark-to-published gate
- Add Resend delivery email and deployed production domain
AI components
- Anthropic API classification of tool name, description and JSON schema into read/write/destructive/external-network with confidence
- Optional second pass to guess the data source behind each tool
Pricing and monetization
Card
Entry offerFor Solo founder or small SaaS with one agent shipping now
- Published card without watermark
- JSON manifest download
- PDF/DOCX answer pack
- One project, 90-day hosting
Review-Ready
For Team actively inside a customer security review
- Everything in Card
- scopecard.yaml lockfile plus CLI diff
- Three re-scans on push
- Custom logo on the card page
Studio
For Small dev agency shipping agents for several clients
- Up to five project cards
- Answer packs per client
- Lockfile plus CLI diff
- Priority re-scan
Model: One-time per project, three tiers, with upgrade and re-publish events as the repeat path.
First revenue offer: $99 to publish your own preview card without the watermark, including the PDF answer pack, paid by card today.
Expected ACV: $99
Upsell path: Upgrade to Review-Ready at $249 when the first deploy changes the tool surface, and to Studio at $499 when an agency needs cards for multiple clients.
First-customer acquisition
GitHub code search for repos containing "mcpServers" and tool-schema files, then one personal email per maintainer containing a watermarked preview card of their own repo.
Where to find them
- GitHub code search "mcpServers" filtered to repos pushed in the last 90 days
- Repos containing claude_desktop_config.json or .cursor/mcp.json
- Repos with tool/function JSON schemas plus an AI-agent README
- One MCP or AI-agent developer community thread where worked examples are welcome
- Small dev agencies publishing agent case studies
Exact outreach motions
- Generate a watermarked preview card for each repo before emailing anyone
- Email the maintainer with their tool count and destructive count as the subject hook
- Follow up once at 72 hours restating the destructive tool count
- Post one worked preview card in a developer community asking for classification feedback
- Email every $99 buyer the $249 upgrade once the lockfile ships
- Ask two buyers for a referral to the reviewer who sent the questionnaire
Message template
Hi {name} - I scanned {repo} and it exposes {N} agent tools, {M} of which write or delete. Here is the capability card at {url} (watermarked, nothing published). It is roughly what your next security reviewer will ask for. $99 publishes it without the watermark and includes a PDF answer pack mapped to the standard questions. Want me to unlock it?30-day execution roadmap
Week 1 · Build & validate · 7 days
Infrastructure and schema
- Create repo and scaffold Next.js with TypeScript on Vercel
- Create Supabase project and tables for projects, scans, tools, cards, purchases
- Create Stripe test prices for $99, $249 and $499
Deliverable: Deployed skeleton with DB schema and Stripe test prices
Repo to tool inventory
- Add GitHub OAuth read access via Octokit
- Add zip-upload fallback path
- Detect mcp.json, claude_desktop_config.json and .cursor/mcp.json plus schema files
- Normalise into a tool inventory with an unclassified bucket
Deliverable: Ingest returning a normalised tool inventory for a test repo
Tool labels with confidence
- Write the Anthropic classification prompt with JSON schema validation and one retry
- Cache classification by commit SHA
- Add manual label override in the UI
Deliverable: Classified inventory for five test repos
Shareable artifact
- Build the hosted card page at /c/{slug} with commit SHA, hash and tool table
- Add JSON manifest endpoint
- Serialise scopecard.yaml from the same record
Deliverable: Live card URL and manifest for a real repo
Sellable MVP live
- Wire Stripe Checkout with a watermarked preview until paid
- Generate the PDF/DOCX answer pack from a 15-question template
- Add Resend delivery email
- Deploy to the production domain
Deliverable: MVP in production with a working checkout
Parser accuracy
- Scan 10 real public MCP repos and log failures
- Add YAML and JSONC fallbacks
- Fix the top five parser gaps
- Manually audit 20 destructive labels
Deliverable: Accuracy log plus fixed parser
Landing page and payment account
- Ship a landing page with one public demo card
- Write pricing and a limitations FAQ
- Run one real card purchase end to end then refund it
- Connect the payment account to TrustMRR
Deliverable: Live landing page, verified checkout, TrustMRR connected
Week 2 · Launch & first sale · 7 days
Build the outreach sheet
- Run GitHub code search for "mcpServers" pushed in the last 90 days
- Filter to repos with parseable tool schemas
- Record maintainer email or contact route for 200 repos
Deliverable: 200-row outreach sheet
Generate evidence per prospect
- Batch-generate watermarked preview cards for the first 80 repos
- Record tool count and destructive count per repo
- Spot-check 10 previews for obvious misclassification
Deliverable: 80 preview URLs with per-repo stats
First 60 sends
- Send 60 personalised emails linking each preview card
- Log sends and opens in the sheet
- Monitor the reply inbox hourly
Deliverable: 60 sends logged
Volume plus community
- Send 60 more personalised emails
- Post one worked preview card in an MCP developer community for feedback
Deliverable: 120 cumulative sends, one community post
Close
- Send 60 more emails
- Run the 72-hour follow-up on wave 1 with the destructive count
- Answer every reply within two hours
Deliverable: About 180 sends and the first paid publish target
Reply to paid
- Handle classifier objections one-to-one with offers to override labels
- Offer the $99 Card directly to anyone who replied
- Refund instantly on request and log the reason
Deliverable: Reply-to-sale conversion log
Upgrade surface
- Ship scopecard.yaml lockfile and CLI diff
- Add re-scan on push webhook for Review-Ready
- Apply first buyer feedback to the answer pack
Deliverable: Lockfile and webhook live
Week 3 · Repeat & systematize · 7 days
Worked example
- Publish an anonymised before/after answer pack with permission
- Add a 'what the reviewer asked' section to the landing page
Deliverable: One public worked example
Message test
- Send 60 emails to a fresh repo batch
- Test a subject line built around the questionnaire deadline
- Log open rates by variant
Deliverable: Subject-line variant data
$249 tier
- Email every $99 buyer the Review-Ready upgrade offer
- Add an upgrade checkout link inside the card dashboard
Deliverable: Upgrade SKU live and offered
Classification accuracy
- Post previews in two AI-agent subreddits and one Discord
- Reply to every classification critique
- Fix the two most-cited label errors
Deliverable: Three posts and a feedback log
Objection removal
- Send 60 emails
- Follow up wave 2
- Fix the three most common objections on the pricing page
Deliverable: 240 cumulative sends
Edge cases and cost
- Test monorepo, YAML, JSONC and private-repo scans
- Add backoff on Anthropic rate limits
- Cap scan cost per card and log token spend
Deliverable: Robustness test log and cost cap
Go or no-go
- Follow up non-openers with their destructive-tool count
- Review conversion from 240 sends
- Decide whether to continue GitHub outreach or pivot to agencies
Deliverable: Written channel decision
Week 4 · Scale what works · 7 days
Second repo shape
- Search GitHub for repos with OpenAPI tool schemas and agent READMEs
- Add 100 rows to the sheet
- Generate 50 preview cards for the new batch
Deliverable: 100 new rows, 50 new previews
Volume plus content
- Send 60 emails to the new batch
- Publish one post showing card anatomy and the hash stamp
Deliverable: 300 cumulative sends
Studio tier
- Email 20 small agencies shipping agents for clients
- Enable a five-project checkout flow
Deliverable: Studio offer sent and multi-project flow working
Warm viewers
- Email every preview viewer who has not bought
- Offer a 20-minute card walkthrough in exchange for a paid publish
Deliverable: Conversion log for warm viewers
Cut refund risk
- Publish a methodology page explaining classification and limits
- Add an explicit 'no runtime enforcement' line to the card and FAQ
Deliverable: Methodology page published
Referrals
- Send 60 emails
- Follow up all unopened threads
- Ask two buyers for a referral to their reviewer contact
Deliverable: 360 cumulative sends and referral asks
Ledger accuracy
- Reconcile Stripe charges against TrustMRR records
- Confirm all payments landed inside the contest window
- Issue any pending refunds and verify cost per card under $0.25
Deliverable: Reconciled ledger
Days 29-30 · Consolidate · 2 days
Last conversions
- Send the final 40 emails with 48-hour framing
- Re-offer $249 to every buyer
- Publish one post answering 'what can your agent access'
Deliverable: Final outreach wave and content
Numbers and learning
- Confirm all payments settled inside the window
- Export card, manifest and ledger backups
- Write a one-page post-mortem on channel conversion
Deliverable: Final numbers and post-mortem
Revenue and profit scenarios
How the numbers are defined
Hackathon profit — Collected revenue − refunds − chargebacks − payment processing fees − paid marketing spend.
Business profit — Hackathon profit − operating costs (hosting, domains, tooling, AI/API delivery costs).
The contest ranks entrants on hackathon profit only. Hosting, tooling and AI/API delivery costs are not deducted by the organisers, but each sale still has to cover them — so we show both. Definitions verified on 2026-11-15 against the official rules.
Stated assumptions
- Conservative: 3 paid publishes from roughly 180 personalised sends.
- Base: 10 paid publishes across ~360 sends, a 2.8% conversion that has not been observed.
- Upside: 25 paid publishes driven by the agency batch and Review-Ready upgrades.
- ARPU held at $99 across all scenarios because the $249 upsell is treated as upside not base.
- Payment processing assumed at roughly 4% of collected revenue (2.9% + $0.30 plus platform fees).
- No paid ads in conservative or base; upside includes $150 of sponsored community and newsletter slots, organic outreach stays the primary channel.
- Operating costs cover Vercel, Supabase, domain and Anthropic classification at under $0.25 per card.
- Upside assumes one refund of $50 and no chargebacks.
Conservative
3 customers- Refunds
- $0
- Chargebacks
- $0
- Payment fees
- $12
- Paid marketing
- $0
- Hackathon profit
- $285
- Operating costs
- $25
- Business profit
- $260
Base case
10 customers- Refunds
- $0
- Chargebacks
- $0
- Payment fees
- $39
- Paid marketing
- $0
- Hackathon profit
- $951
- Operating costs
- $45
- Business profit
- $906
Optimistic
25 customers- Refunds
- $50
- Chargebacks
- $0
- Payment fees
- $97
- Paid marketing
- $150
- Hackathon profit
- $2,178
- Operating costs
- $90
- Business profit
- $2,088
Model your own numbers
Runs in your browser · nothing is saved- Collected revenue
- $990
- Payment fees (3.9%)
- $39
- Paid marketing
- $0
- Hackathon profit
- $951
- Business profit
- $906
Hypothetical arithmetic, not a forecast. The fee rate is a flat 3.9% assumption, refunds and chargebacks are left at zero here, and none of this is tracked data.
These are projections built from the stated assumptions, not verified revenue, customers or traction.
Main risks and validation experiments
Risk 1Statically parsed cards miss dynamically registered tools, so the published artifact is incomplete.
Ship an explicit unclassified bucket on every card and never label a card verified safe.
Scan 20 real public MCP repos by day 6 and count unclassified tools per repo.
Abandon the static-scan promise if more than 30% of tools are unclassified on half the sample.
Risk 2Cold GitHub outreach replies well under 3%, making any sales target unrealistic.
Lead every email with the recipient's own destructive-tool count and one 72-hour follow-up.
60 sends by day 10, measure replies by day 12 and open rate by variant on day 16.
Pivot to the agency and warm-replier angle if replies are under three from 180 sends.
Risk 3Buyers expect runtime enforcement and refund a document.
State 'static attestation, no runtime interception' on the card, pricing page and methodology page.
Track refund reasons from day 12 and publish the methodology page on day 26.
Reposition if refunds exceed 10% of sales or two buyers cite enforcement in writing.
Risk 4A destructive tool is mislabelled as safe, embarrassing the customer in front of their prospect.
Confidence flags, manual override and a default-to-unknown rule for ambiguous schemas.
Manually audit 20 destructive labels on day 6 and 10 more on day 20.
Suspend publishing for any repo whose audit error rate exceeds one in ten.
Risk 5Anthropic rate limits or token cost spikes slow scans during an outreach wave.
Batch one call per scan, cache by commit SHA, add backoff and a per-card cost cap.
Run a 20-scan burst on day 20 and measure latency and spend.
Move to a cheaper classification model if cost per card exceeds $0.25.
Hour-by-hour: the first 24 hours
Hours 1-3
- Create the ScopeCard repo and scaffold Next.js with TypeScript
- Create the Supabase schema for projects, scans, tools, cards and purchases
- Create Stripe test prices at $99, $249 and $499
- Deploy the empty app to Vercel
Output: Deployed skeleton with database schema and Stripe test prices
Hours 3-6
- Wire GitHub OAuth read access with Octokit
- Implement config detection for mcp.json, claude_desktop_config.json and .cursor/mcp.json
- Run detection against two known public repos
Output: Ingest returning a raw file and tool list for two real repos
Hours 6-10
- Write the Anthropic classification prompt with strict JSON output
- Validate and retry once on malformed output
- Persist the classified inventory keyed by commit SHA
Output: Classified tool inventory stored for both test repos
Hours 10-14
- Build the card page at /c/{slug}
- Render commit SHA, content hash and the tool table
- Add the unclassified bucket
Output: Live, shareable card URL for a real repo
Hours 14-18
- Wire Stripe Checkout
- Add the watermark-to-published gate
- Generate the PDF answer pack and send it by Resend
- Deploy to production
Output: End-to-end purchase flow that works on a real card
Hours 18-24
- Run the full flow on one unfamiliar repo
- Fix the two worst failures in the parser
- Generate the first three preview cards for real public repos
Output: Three preview cards queued for day-10 outreach and a short failure log
Full 30-day launch report
Read the full generated report
1. Product opportunity and positioning
An AI agent is a set of tools with permissions. Buyers of small B2B software have started asking for proof of that surface inside security reviews: which data can the agent reach, which actions can it take, which credentials does it hold, where does a human approve. The founder answering today re-reads tool schemas and mcp.json by hand, writes fresh prose per prospect, and cannot show the answer matches the deployed commit.
ScopeCard converts the repo into the answer. Connect a GitHub repo or upload a zip; ScopeCard parses MCP config files and tool/function schemas, classifies each tool as read, write, destructive or external-network with a confidence flag, and publishes a hosted capability card stamped with the scanned commit SHA and a content hash. The same scan emits a JSON manifest, a scopecard.yaml lockfile and a PDF/DOCX answer pack mapped to the roughly fifteen questions agent reviews actually ask.
Positioning is narrow on purpose: not a runtime proxy, not observability, not an identity fabric. The buyer is the technical founder at a 1-20 person B2B SaaS with one shipped agent and one prospect blocking the deal. The trigger is a dated questionnaire, the approval is a single card payment, and the deliverable is a document generated from code in one session, so value lands before the invoice clears.
JEV scores speed to first revenue highest at 7.72 normalized and market demand lowest at 3.98. That is the strategy in one line: the product is fast to sell if the buyer exists, so demand is a hypothesis to test by day 12, not an assumption to celebrate on a landing page.
2. Smallest sellable MVP
Five things must exist before the first dollar. Repo ingest by GitHub OAuth or zip upload. A classifier that labels every tool read, write, destructive or external-network with confidence, plus a UI override. A hosted card at a stable slug showing commit SHA and content hash. A paid unlock that removes the watermark and releases the manifest and PDF/DOCX answer pack. Stripe Checkout. Everything else is an upgrade shipped after day 12: lockfile CLI diff, re-scan webhooks, custom logo, multi-project.
The completeness test is single: can a stranger open a preview URL, recognise their own tool list, pay $99 and forward the PDF to their prospect within ten minutes without talking to the founder. Runtime interception never ships in this window.
3. Technical implementation plan
Next.js on Vercel, Supabase Postgres, Octokit, Anthropic API, react-pdf, Stripe Checkout, Resend. Ingest is a serverless route that pulls the tree, matches known config filenames plus any JSON or YAML file containing tool or function schemas, and normalises to an inventory keyed by name. Classification is one batched Anthropic call per scan with strict JSON output and one retry, cached by commit SHA, so re-renders cost nothing.
The card page, manifest and lockfile all render from one stored record plus its content hash, so they cannot disagree. Payment flips a published flag scoped to one project for 90 days. The hardest technical risk is parser recall on monorepos and JSONC; mitigate with an explicit "tools we could not classify" bucket instead of silent drops.
4. Pricing and monetization
Three one-time tiers: Card $99, Review-Ready $249, Studio $499. One-time fits a single review, but the repeat path is real: every new prospect questionnaire and every deploy that changes the tool surface creates a re-publish or upgrade event, and Studio turns one agency into five projects.
Processing is roughly 4% and classification under $0.25 per card, so margin holds above 90% at $99. Do not discount below $99. Instead trade a public worked example to the first three buyers for permission to describe the questionnaire they answered.
5. First-customer acquisition
One channel: GitHub code search for "mcpServers" and tool-schema files, filtered to repos pushed in the last 90 days, then one personal email per maintainer containing a watermarked preview card of their own public repo. The email leads with their numbers, not a pitch.
Target 360 sends across seven waves with one 72-hour follow-up each, and a hard read at day 21: fewer than three paid publishes from 240 sends means the channel or the offer is wrong, and effort moves to warm repliers and the agency angle instead of more volume on a dead list.
6. Thirty-day execution roadmap
Days 1-5 build to a live checkout and day 7 verifies a real purchase with the payment account connected. Days 8-12 generate 200 preview cards and send 180 emails, with day 12 as the first-sale target. Days 13-18 convert replies, ship the lockfile and re-scan upgrades, publish one anonymised worked example, and post in developer communities asking specifically about classification accuracy. Days 19-21 run a robustness pass and the channel review. Days 22-27 add an agency batch, a methodology page to cut refund risk, and two more waves. Days 28-30 reconcile Stripe against TrustMRR, run the final push and write the post-mortem. Every day has one deliverable and no day carries more than four tasks.
7. Revenue and profit scenarios
Projections only, derived from stated assumptions, not verified revenue. Conservative: 3 sales at $99 is $297 gross, $12 processing, $25 operating, $285 hackathon profit, $260 business profit. Base: 10 sales is $990 gross, $39 processing, $45 operating, $951 hackathon profit, $906 business profit. Upside: 25 sales is $2,475 gross, $50 refunds, $97 processing, $150 sponsored community slots, $90 operating, $2,178 hackathon profit, $2,088 business profit.
Base implies about 10 paid publishes from 360 personalised sends, roughly 2.8% conversion, which has not been observed. The contest ranks hackathon profit, which ignores hosting and API delivery cost; the business column is what the founder keeps.
8. Main risks and validation experiments
Parser recall: if cards miss dynamically registered tools the artifact is worse than useless. Scan 20 real public MCP repos by day 6 and count unclassified entries; widen the parser or narrow the promise if more than 30% of tools are unclassified on half the sample.
Demand: cold GitHub reply rates may sit under 3%. Run 60 sends by day 10 and measure opens and replies by day 12; under three replies from 180 sends, re-cut the offer around the questionnaire deadline or move to agencies.
Willingness to pay for a document instead of software: watch refund requests citing expected enforcement. Put the limitation on the card page and FAQ by day 26 and track refund rate; above 10% of sales the positioning is wrong. Secondary risks: mislabelling a destructive tool as safe, mitigated by confidence flags and manual override and by never calling a card verified safe; and Anthropic rate limits during an outreach push, mitigated by backoff and per-commit caching.
9. First 24-hour action plan
Hours 1-3: repo, Supabase schema, Stripe test prices, Vercel skeleton deployed. Hours 3-6: GitHub OAuth and file detection working against two known public repos. Hours 6-10: classification prompt producing validated JSON, inventory stored per commit. Hours 10-14: card page rendering commit SHA, hash and tool table. Hours 14-18: Stripe Checkout, watermark gate and PDF export, deployed. Hours 18-24: run the flow end to end on one unfamiliar repo, fix the two worst failures, and generate the first three preview cards for real repos to seed day-10 outreach.
This brief was generated from public data about established startups, then narrowed and scored for a solo 30-day launch. Revenue and profit figures are scenario models, not verified results, guarantees or evidence of existing customers. Validate demand yourself before you spend a month on it — and check the contest rules for what counts toward your score.
Related ideas
All ideas →Retainer Proof
Generates a branded, client-ready monthly care-report PDF for small WordPress agencies from free public signals, with nothing installed on the client site.
Inbox Passport
Inbox Passport fingerprints a client domain's real email sending stack, generates the exact SPF, DKIM and DMARC records to paste, verifies them over 48 hours, and issues a timestamped white-label PDF certificate the agency can hand to its client.
Idle Hours GPU Waste Autopsy
Upload your cloud/GPU billing export and get a dollar-quantified PDF autopsy of exactly where your GPU spend is being wasted, ranked by recoverable dollars per month.