FiBuKI plugin for ChatGPT, Codex and Claude

Pre-accounting for Austrian one-person businesses, from inside an AI assistant: set up, import a bank CSV, find and match the invoices for a month, hand the Steuerberater clean data.

This is a skills-only plugin on top of FiBuKI's MCP server. There is no FiBuKI logic in it: FiBuKI reads the documents, scores the matches and stores everything; the skills drive the loop and ask the user once per batch. Plan and decisions: handoffs/2026-10-02-openai-plugin.md.

plugin.json  mcp.json          portable Agent Plugins package (schemas validated)
skills/
  fibuki-onboarding/           identity, bank data, mailbox, first Belege
  fibuki-belege/               match a month, hunt missing invoices, close out
    references/invoice-hunting.md
  fibuki-bank-csv/             bank CSV -> Transactions
    references/austrian-bank-csvs.md
  (each: SKILL.md, agents/openai.yaml, evaluations/*.json)
scripts/
  fibuki-csv.mjs               GENERATED bundle of lib/import + src/csv-cli.ts
  fibuki-upload.mjs            local PDFs/images -> upload_file
src/csv-cli.ts  build.mjs      source and bundler for fibuki-csv.mjs

Use it with Codex

Add FiBuKI's MCP server to ~/.codex/config.toml. No key needed: FiBuKI is an OAuth server, and Codex discovers it from the 401 challenge, registers itself and signs the user in.
```
[mcp_servers.fibuki]
url = "https://fibuki.com/api/mcp/sse"
```
Sign in: codex mcp login fibuki. A FiBuKI window opens; sign in or create an account (a new user is asked who they are first), then choose Allow. The connection shows up under Settings, Integrations as a connected app and can be revoked there.

Alternative, with an API key (tools without OAuth, or a self-host without a public web origin): create one with npx @fibukiapp/cli auth --format env, export FIBUKI_API_KEY=fk_..., and use bearer_token_env_var = "FIBUKI_API_KEY" in the same block instead of logging in.

The scripts still need an API key. fibuki-csv.mjs --import and fibuki-upload.mjs call FiBuKI directly rather than through the MCP connection, so they read FIBUKI_API_KEY whichever way the MCP server is connected.

Install the plugin from a local marketplace: put this folder at <repo>/plugins/fibuki and add to <repo>/.agents/plugins/marketplace.json:

{
  "name": "local-repo",
  "plugins": [
    {
      "name": "fibuki",
      "source": { "source": "local", "path": "./plugins/fibuki" },
      "policy": { "installation": "AVAILABLE", "authentication": "ON_INSTALL" },
      "category": "Finance"
    }
  ]
}

Try: "Set me up with FiBuKI", "Mach meinen letzten Monat fertig", "Import my bank CSV".

Self-host: use your own URL in step 2 and set FIBUKI_BASE_URL for the scripts.

Not yet verified in a real Codex or ChatGPT session. Built and tested against the manifest schemas, the MCP server (a real SDK client over HTTP) and the scripts (fixtures and a fake API). First thing to check in Codex: that the bundled mcp.json server and your config.toml entry do not collide; if they do, drop one.

The skills use the connected mail and file apps of whichever assistant runs them (Gmail, Outlook, Google Drive, OneDrive, Dropbox). The plugin does not require them: without any, it still imports, matches what FiBuKI already has, and lists what is missing.

The scripts

# what is in this file?
node scripts/fibuki-csv.mjs analyze export.csv [--skip N]

# dry run: converts, writes chunk files, prints totals and a preview. Sends nothing.
node scripts/fibuki-csv.mjs convert export.csv --date Buchungsdatum --amount Betrag \
  --name Buchungstext --partner Partnername --iban "Partner IBAN" --balance Saldo \
  --after 2026-08-31

# send it (FIBUKI_API_KEY, import_transactions in chunks under one import id, waits out
# rate limits). FiBuKI skips lines it already has; the result says how many (alreadyImported).
node scripts/fibuki-csv.mjs convert export.csv ... --import --source <sourceId>

# receipts and invoices
node scripts/fibuki-upload.mjs a1-rechnung.pdf bon.jpg

fibuki-csv.mjs is built from FiBuKI's own lib/import parsers, so it reads dates and amounts exactly as the web import does. On top of that it checks the result against every row: day/month order, the decimal mark, and (with --balance) that the running balance adds up. After changing lib/import or src/csv-cli.ts:

node integrations/openai-plugin/build.mjs         # rebuild the committed bundle
node integrations/openai-plugin/build.mjs --check # CI: fails when it is stale
node --test tests/openai-plugin-csv.test.mjs tests/openai-plugin-upload.test.mjs

Known gaps

No shell, no CSV import. Plain ChatGPT has no shell, so the CSV skill sends the user to the web import. A server tool that reuses the same parsers would close this (phase 4).
No tool lists Postfächer, so the skills ask the user.
Bank layouts are modelled, not collected. Add anonymised real exports to tests/fixtures/openai-plugin and extend references/austrian-bank-csvs.md.

Widgets

Three small in-chat views ship with the MCP server itself (functions/src/mcp-api/widgets.ts), not with this package: an onboarding checklist (get_onboarding_status), a progress board (get_period_status) and a match review (list_pending_matches). Hosts that support MCP Apps / the ChatGPT Apps SDK draw them next to the tool result; every other client just sees the tool's JSON, so the skills never depend on a widget. They only call tools that already exist (connect_file_to_transaction, dismiss_transaction_suggestion, auto_connect_file_suggestions, upload_file). In ChatGPT the progress board also offers a file picker that hands the chosen files to upload_file through its file argument.