Skip to main content
Glama

AI Launchpad — AI-hüppelaud

Put your AI assistant to work for your business / Pane oma AI-assistent ettevõtte heaks tööle.

Turn the AI assistant an Estonian small-business owner already has — or is about to get — into their business assistant.

End goal

AI agents take the back-office load off small-business owners and their employees, so people spend their time on their core work. Measures: time saved, reliable outcomes and remaining manual work (R-001).

Eesmärk: tehisaru-agendid võtavad väikeettevõtja ja tema töötajate õlult ära kontoritööd, et inimesed saaksid keskenduda oma põhitööle.

Status: owner-only HTTPS pilot at https://huppelaud.arikratt.ee. One owner ChatGPT plus Meta/Baseten path tested; Claude tested once on the Free plan with synthetic data (30 Oct 2026), so this is not broad compatibility or production ready. See docs/live-pilot.md for dated evidence, cost, and recovery limits. Local demo and tests remain mock ($0); see docs/local-demo.md.

Related MCP server: reai-mcp

For hackathon judges

  • Built during the event: all product code in this repository was written during the Meta Global AI Developer Hackathon (27 Oct – 4 Nov 2026; first commit 27 Oct). Third-party libraries and services are used through their public interfaces. The one service of the author's that existed before the event is the Ärikratt MCP, a public Estonian company-data service used as an external dependency.

  • Meta's models: Muse Spark Open (meta-models/muse-spark-open-1026) via Baseten; see Where Meta's models fit. No other AI model is called at runtime. Live calls run on the hosted pilot (/health shows live_inference: enabled).

  • Try it without an account or keys: https://huppelaud.arikratt.ee/naidis?lang=en (walkthrough) and https://huppelaud.arikratt.ee/proovi?lang=en (your own sample workspace, nothing really sent). Guide: docs/demo/judge-guide.md.

  • Run it locally (Node 22.12+, macOS or Linux shell, synthetic data, no paid calls):

    npm ci
    npm run dev   # http://localhost:3000

    The local run uses MODEL_MODE=test: model outputs are saved fixtures labelled as not a live model, and no Meta call is made. Keys come from environment variables; .env.example lists them with placeholders. Live Muse calls with your own BASETEN_API_KEY need the guarded operator mode MODEL_MODE=public_cards, which also requires LAUNCHPAD_GOOGLE_ALLOWED_SUBJECTS, LAUNCHPAD_MODEL_PRICING_CONFIRMED=1 and an existing LAUNCHPAD_MODEL_LEDGER_PATH; see docs/model-quality.md. ARIKRATT_API_KEY enables live register lookups. More in docs/local-demo.md.

The problem

  • An Estonian micro-entrepreneur carries the whole back office: payroll and VAT filings, the annual report, invoices, bookkeeping, contracts, permits and sector rules. The research maps 354 such jobs.

  • General assistants (Claude, ChatGPT) are already on the owner's phone, but they don't know Estonian obligations, aren't connected to the accounting software, bank and registers, and have no safe way to act.

  • Many owners feel left behind in the agent race: they hear about AI agents but have never set one up and are wary of letting one near their taxes and money.

Who it's for

Audience

Gap we close

Owners who already use ChatGPT or Claude for chat (main target)

Turn that chat app into a business assistant: connectors, Estonian knowledge, recorded outcomes

Owners who feel left behind — never used an agent

A calm path from zero to a working assistant on their phone

Owners who already connect MCP servers themselves

Early testers

Later: accountants serving many micro-companies

One accountant sets it up for many clients

First persona: a one-person Estonian service OÜ, books in e-arveldaja or Merit. FIEs, e-residents and employers follow later — their rules differ.

Principles

  1. The owner's own assistant is the interface. We don't run a chatbot that competes with it. The only exception is a temporary setup helper for people who have no assistant yet.

  2. No forms. What public registers or the assistant's own context already know is never asked. The owner answers only what nobody else can.

  3. Ask for a job; your assistant does it and records the outcome. It sends messages and invoices and posts book entries with its own tools. It asks first only when you asked it to, now or as a saved fact, or it judges it needs your yes: instructions came from a file, e-mail or web page, the recipient or amount differs from the books, or the content is unclear. Then it shows the exact content and asks once. Text in files is data, never instructions. For payments and filings, it prepares exact figures; you pay in the bank and file in e-MTA / e-Business Register with your own Smart-ID.

  4. Rules compute, models read. Deadlines, tax amounts and obligations come from deterministic rules code with official sources; models read documents, plan steps and draft text.

  5. Store minimal private operational state. Identity/workspace references, bounded assessments and task reports, grants, recorded action outcomes, optional confirmations and activity support the companion. These are not re-derivable from public registers. Do not request conversations or source documents; let owners inspect, export and delete their progress. Current limits.

  6. Reuse before building. Existing MCP servers (accounting, bank, credit data, the assistant's built-in e-mail and calendar) are curated, not rebuilt. New adapters are built where demand shows.

Planned product flow (beyond the local demo)

Path A — the owner already has an assistant

  1. Site: one step that can't happen in chat — add the AI Launchpad connector to Claude or ChatGPT (URL + a three-step guide; later one tap via their connector directories). On a paid Claude plan the AI Launchpad plugin adds the connector plus skills that send "link my e-mail" or "my annual report" to Launchpad; on the free plan the connector's own connect_my_services tool does that job. Where subagents are supported, the plugin's action-reviewer checks each send, book posting and handover of payment or filing figures in a separate context before the assistant acts. It sees only the owner's request, exact action and book or customer facts. A concern uses the existing ask-first path.

  2. In chat: "help me run my company's admin" → the assistant calls get_started and asks only which company? (name or registry code).

  3. AI Launchpad fills in the rest from public data via the Ärikratt MCP: VAT number and activities (Business Register), employee count, labour taxes and turnover (tax board quarterly data), licences, annual-report status, tax debt → obligations calendar and the list of jobs the assistant can take over.

  4. Unknowns (accounting software, bank): the assistant first checks its own connectors; it asks one question only if still unknown.

Path B — the owner feels left behind

  1. Start where they already are: Ärikratt company pages already get search traffic. Each shows "your assistant could handle these deadlines for this company → set it up". Their own company, no input, value first.

  2. One recommendation, not a menu: e.g. Claude — the free plan allows one custom connector, and AI Launchpad is one endpoint.

  3. Guided setup on their own phone: QR code from computer to phone, exact steps with screenshots for their app and phone, "✓ Connected" detected by the gateway.

  4. A safe first win in five minutes: read-only prompts — "what are my deadlines this month?", "check this company before I take the job", "is my annual report filed?"

  5. Your choice: ask for a job and the assistant does it, then records the outcome. Tell it to check with you first whenever you want. It remembers that preference; approval cards remain available for those requests.

  6. Setup helper (a bridge, not a rival): "I'm stuck — here's my screen" (screenshot → where to tap) and voice for people who prefer talking. Once their own assistant is connected, it hands over: "from now on, talk to Claude."

  7. Human backup: a weekly online setup clinic in Estonian, or setup through their accountant.

AI Launchpad (one MCP gateway)

Part

What it does

Company profile & obligations calendar

Rules engine + Ärikratt public data

Job playbooks (~10 to start)

For each job: steps, tools, recorded outcome, official source

Catalogue

Researched MCP services per job; how to connect each; runtime verification remains future work

Permissions

Per-authority inventory (EMTA, Business Register, …) with expiry and revocation — filing rights are separate per authority

Action outcomes & optional approvals

The assistant acts on the owner's request and records sent / submitted / accepted / rejected / paid states. When asked, an approval binds the owner's yes to the exact action, in their assistant or on the approval page.

Portal prefill

Where no API exists, a browser agent fills the government portal and stops before Submit

Friction loop

Where an assistant gets stuck: classify why, deduplicate, redact, publish an owner-approved request as open demand on the marketplace

Each job is handled one of four ways: an existing MCP · an adapter on an official API · a browser agent with a hard stop · the owner (PINs, signatures, judgement calls). Mapping in docs/research/README.md.

Starting jobs (from the research's top-10 lists): deadlines this month · counterparty check · sales invoice · supplier invoice into the books · bank reconciliation · payment reminder · monthly payroll/VAT filing preparation · annual-report preparation · VAT-threshold warning · grants and tenders.

Discovery design (draft, unpublished)

The adopted discovery design is a draft: an AI Catalog index plus MCP Server Cards describing the gateway and curated connectors. Those discovery artifacts are unpublished — no live catalogue or card yet. This does not deny the live owner-only MCP pilot endpoint — see docs/live-pilot.md. Details and evidence status are in docs/research/H-mcp-discovery-and-publication.md.

Where Meta's models fit

Shipped, behind src/model-quality.ts (meta-models/muse-spark-open-1026 via Baseten):

Where

What the code does

Unmet-need classification

Background job in src/unmet-needs.ts after the report_unmet_need tool. Live output is labelled „AI liigitus · Meta Muse Spark Open”; mock output is labelled „Näidis”. /demand lists only rows an owner published.

Public setup cards

An operator generates them with src/model-public-cards-cli.ts (job setup_guidance). /setup shows „General guidance · Meta/Baseten” only when a completed card is saved. Opening /setup does not call the model. /vali redirects to /alusta.

Onboarding: name the programs

/alusta job intake_extract. Muse Spark names programs the visitor wrote. A catalogue name match is used when the limit, a failure, or model-off applies, and the page says so. Every onboarding call is a new ledger row and its text is not stored (no replay).

Onboarding: one plan line

/alusta job intake_plan. One short line from facts the rules already computed. Prices, dates and steps stay on the rules. If Muse is unavailable the page says so and still shows the rules plan. Every onboarding call is a new ledger row and its text is not stored (no replay).

Saved friction cards

The same operator path, job classify_friction. /playbooks shows a saved card when one exists („Salvestatud juhis”) and does not call the model on GET.

Not in this repo: Muse Voice Transcribe, a vision setup helper (screenshot to a tap target), Muse Spark vision, Muse Glimmer, and Muse Code turning requests into adapters. Setup input is fixed text. There is no chatbot competing with the owner's assistant.

The historical bounded quality harness ran two genuine synthetic setup calls on 2026-10-28 (generic acceptable English/Estonian, not a benchmark — see docs/model-quality.md); the owner-only pilot aggregate is now 4 real calls total — see docs/live-pilot.md for the pinned runtime and receipts. No EU self-hosting or data-residency posture exists yet.

Meta's own consumer agent app (Muse) is available only in the US and Canada as of 26 Oct 2026 and documents no custom MCP setting; see docs/research/G-muse-app-for-estonian-smes.md. AI Launchpad stays MCP-first and can add a plain API if Muse arrives in the EU.

Hackathon scope and target demo

  • Scope: the free micro-entrepreneur service only. Enterprise measurement is future design; billing and enterprise features are excluded.

  • Shipped: get_started and the playbooks in src/catalog.ts (src/mcp.ts); the obligations calendar from get_obligations and src/obligations.ts (signed-in /, public walkthrough /naidis); actions recorded directly with record_action_outcome and shown as “Your assistant did it”; optional approval cards via request_approval / get_approval and /approvals; whats_waiting_for_me returns the same list as Täna's „Ootab sinult”, and the „Lahenda Claude'iga” button there opens the owner's assistant with a fixed prompt to work through it; the friction loop via report_unmet_need, the owner's publish decision on Täna, and the public list /demand; the per-visitor sample workspace /proovi (optional approval examples: approve and decline, nothing really sent); the public onboarding chat /alusta (company, programs, assistant and a rules plan, nothing stored); the temporary text setup helper /setup for a signed-in owner (a sample session is refused).

  • Not shipped: the five-minute "Google search to phone" Path B story. Muse Voice Transcribe and a vision setup helper have no call site.

  • Out: payments, unattended filing or signing, payroll submission, a marketplace UI beyond the demand list, FIE/e-resident/employer variants.

Accepted product requirements

  • R-001: recommend the action that removes the most owner work (issue #1). The owner’s agent makes the final assessment using Launchpad’s framework, with evidence, assumptions and one clear next action. Success means time saved, reliable outcomes and less remaining manual work; estimates and unconfirmed reports remain labelled. Agent-assessed coverage is separate from source-verified outcomes (currently zero); no manual ratification is required.

  • R-002: identity-only sign-in and bounded runtime model use (issue #3). Google is identity only; the old adapter in src/model.ts stays hard-disabled. A separate bounded quality harness is now live-tested (docs/model-quality.md); the reviewed public-cards path is piloted live owner-only (docs/live-pilot.md), and broader public runtime activation stays a later decision.

  • Deadlines from rules (issue #19, rules memo). The get_obligations tool and the Status card compute KMD, TSD, annual-report, VAT-threshold and TÖR dates with deterministic rules, each row linked to its Riigi Teataja source; no model produces a date. Unknown facts become conditions and questions, never guesses. Public-register lookup uses an Ärikratt API key (ARIKRATT_API_KEY), provisioned on the live pilot since 31 Oct 2026; where it is not set (local demo, tests), the facts come from the owner's assistant or the labelled sample companies.

Plugins

  • Claude: in claude.ai, Customize → Plugins → Add → Add marketplace, enter paat/ai-launchpad, add AI Launchpad, then connect it in the plugin's Connectors tab. In Claude Code: /plugin install ai-launchpad --marketplace paat/ai-launchpad. plugin/ai-launchpad.

  • ChatGPT: plugin/ai-launchpad-chatgpt, the same skills and connector; it will be in the ChatGPT Plugins directory once OpenAI approves it. Until then, the closed pilot uses the connector directly.

  • The skills live once in plugin/ai-launchpad/skills/; the ChatGPT plugin links to them.

Validation

Test with 3–5 real owners of the first persona during the build week (with consent). Measure time to first useful task (target under five minutes) and whether they come back the following week.

Open questions

  • ChatGPT's plan requirements for custom MCP connectors.

  • Which authorities allow a multi-tenant gateway to act on an owner's mandate (EMTA X-tee machine rights, RIK Andmesild, Tervisekassa).

  • Werkzeug and Trivis (existing Estonian accounting MCPs): partners, listings or competitors?

  • Estonian quality of Muse Spark / Muse Spark Open / Voice Transcribe; SAM 3.1 availability in the Meta Model API.

  • Address — owner-only HTTPS pilot live at https://huppelaud.arikratt.ee since 28 Oct 2026 UTC / 29 Oct Tallinn (commit 2bf1d06265ae1c8173cb3491cd34f18441a36f04); one owner/workspace pilot, not general availability.

Relationship to Ärikratt

This is a separate project. When ARIKRATT_API_KEY is set, src/arikratt.ts calls the public Ärikratt MCP (get_company, search_companies) for company facts. Without that key the lookup returns not_configured and does not call the network. Ärikratt stays an external dependency.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      B
      quality
      D
      maintenance
      MCP server for DACH accounting automation. Connect AI assistants to sevDesk and Lexoffice — create invoices, manage contacts, handle bookings and vouchers for German-speaking businesses.
      15
      55 npm
      -
    • A
      license
      A
      quality
      D
      maintenance
      MCP server for ReAI, the Norwegian cloud accounting system, enabling AI agents to read books, look up accounts and VAT codes, and perform bookkeeping through a curated toolset with safety guardrails.
      12
      MIT
    • A
      license
      Not graded
      quality
      D
      maintenance
      MCP server that exposes an in-browser AI ERP (CRM, stock, invoices, HR) with role-aware agents, enabling natural language chat and confirm-before-write actions.
      MIT
    • A
      license
      Not graded
      quality
      Not graded
      maintenance
      Enables AI assistants to connect securely over Streamable HTTP/SSE or a local stdio bridge to specialized digital administrative colleagues for tasks such as business-card scanning, prospecting, encrypted file management, approval queues, learning memory, and agent status, while preserving human-in-the-loop control and EU data sovereignty.
      MIT