Stackin MCP plugin — NF-e and NFS-e for ChatGPT, Claude Code, Codex and Cursor

---

Ask in plain language — "issue an NFS-e for this service", "why was my last NF-e rejected?", "cancel invoice 1042 and tell me the reason code" — and the model does it through Stackin, against SEFAZ and the municipal NFS-e webservices.

Sixteen tools, nine skills, one hosted server. No SDK to install, no XML to sign, no certificate to juggle in the client.

• 🧾 NF-e (goods) and NFS-e (services) — issue, consult, cancel
• 🔎 Diagnose a rejection by the tax authority's own code, and choose reissue vs. correction
• ✉️ Correction letter (CC-e) against an authorized NF-e
• 📥 Received invoices — read and manifest documents other companies issued against yours
• 🔢 Numbering invalidation for a reserved, never-used NF-e range
• 📄 DANFE / DANFSe PDF straight out of the conversation
• 🔎 Fiscal code lookup — resolve an NCM, CFOP or CEST instead of guessing one, and confirm who a CNPJ belongs to

Table of contents

• Install in Claude Code
• Install in ChatGPT and Codex
• Connect from Cursor or any MCP client
• Authentication: OAuth or API key
• Skills
• What is in this repository
• One plugin, three clients
• MCP Registry
• FAQ
• Maintainer notes

Install in Claude Code

The repository is its own marketplace, so it installs straight from GitHub:

claude plugin marketplace add stackin-io/stackin-mcp-plugin
claude plugin install stackin@stackin

The plugin brings the nine skills and points Claude Code at https://mcp.stackin.io/mcp. There is nothing to configure: the first tool call answers 401 with WWW-Authenticate, and Claude Code follows it — run /mcp in the session, authorize in the browser that opens, and the token is stored for you. The client registers itself through /oauth/register, so no client_id is handed out by anyone.

Prefer a static key? Add the server by hand with a header instead of installing the plugin — see Authentication.

Install in ChatGPT and Codex

The same server is published as a ChatGPT app over Developer Mode. Enable developer mode, register https://mcp.stackin.io/mcp as a connector, and authorize — the full walkthrough, with video, is at docs.stackin.io/mcp.

Connect from Cursor or any MCP client

Any MCP client can use the server directly, without this plugin. Cursor reads .cursor/mcp.json in the project, or ~/.cursor/mcp.json. The interpolation syntax is not the same as Claude Code's: Cursor wants ${env:VAR}.

{
  "mcpServers": {
    "stackin": {
      "url": "https://mcp.stackin.io/mcp",
      "headers": { "Authorization": "Bearer ${env:STACKIN_API_KEY}" }
    }
  }
}

Claude Code, by hand — ${VAR} expands from the environment:

claude mcp add --transport http stackin https://mcp.stackin.io/mcp \
  --header "Authorization: Bearer ${STACKIN_API_KEY}"

Property	Value
Endpoint	https://mcp.stackin.io/mcp
Transport	streamable-http
Auth	OAuth 2.1 with PKCE and dynamic client registration, or a static Bearer key
Environments	Homologation and production, fixed on the key or on the authorized connection

Authentication: OAuth or API key

Two ways in, and they behave differently.

An API key. Generate one at app.stackin.io; it already carries the environment, so there is nothing to choose in the client. Send it as Authorization: Bearer <key>.

OAuth, with no key at all. Leave the header out: a tool call answers 401 with WWW-Authenticate pointing at /.well-known/oauth-protected-resource, which is what makes Claude Code offer to log in when you run /mcp. The server supports dynamic client registration (/oauth/register), so a client that registers itself gets a client_id without anyone being asked. Cursor can also hold static client credentials under an auth key instead of registering dynamically.

Verified on 2026-09-09, up to the browser: the server connects with no header at all, a tool call answers 401 with the resource_metadata challenge, the protected-resource document resolves, /oauth/register returns 201 and /oauth/authorize redirects to the consent screen. The browser leg itself has not been walked from Claude Code or Cursor — ChatGPT completed a real consent, those two have not. If one stops midway, look at the redirect_uri it uses: /oauth/authorize only accepts a URI the client registered.

Skills

A skill exists only where the order of operations, or a fiscal rule, adds something a tool description cannot carry on its own.

Skill	What it carries
issue-invoice	Which document the situation needs, and validating before issuing
consult-invoice	Finding a document, and which identifier each operation takes
cancel-invoice	Confirming before an irreversible act, and when a correction letter is the right tool instead
diagnose-invoice	Reading the tax authority's own rejection code, and choosing between reissue and correction
correct-invoice	Filing a correction letter against an authorized NF-e, and what a CC-e may not change
invalidate-numbering	Declaring an NF-e numbering range reserved and never used
received-invoices	Reading documents other companies issued against this one, and answering them
fiscal-lookup	Resolving an NCM, CFOP or CEST before issuing, instead of guessing one
taxpayer-lookup	Confirming who a tax id belongs to, and why a 404 is not an invalid CNPJ

What is in this repository

This repository is the distribution layer: the manifests, the skills and the assets. The server itself lives at https://mcp.stackin.io/mcp and speaks the Model Context Protocol.

Path	What it is
.codex-plugin/plugin.json	Plugin identity, version, category, legal links
.app.json	Maps the plugin to the registered MCP connection
.agents/plugins/marketplace.json	Local marketplace, for installing before publishing
.claude-plugin/plugin.json	Plugin identity for Claude Code
.claude-plugin/marketplace.json	This repository as a marketplace Claude Code can add
.mcp.json	The hosted MCP server, for Claude Code
server.json	The MCP Registry entry
skills/	The nine workflows that need more than a tool description
assets/	Icon and logo

One plugin, three clients

The repo carries three manifests because three ecosystems look in three places, and each one fails silently when its file is missing:

Client	Looks for	What it reads
ChatGPT / Codex	.codex-plugin/plugin.json + .app.json	the app id and the hosted MCP
Codex marketplace	.agents/plugins/marketplace.json	this repo as a source
Claude Code	.claude-plugin/marketplace.json, .claude-plugin/plugin.json, .mcp.json	the skills and https://mcp.stackin.io/mcp

Adding the folder to Claude Code failed with "No marketplace manifest found … (expected .cursor-plugin/marketplace.json or .claude-plugin/marketplace.json)" until the third row existed. The skills are shared: all three point at skills/, so a skill written once is read by every client.

MCP Registry

server.json describes the hosted server for the official MCP Registry as a remote entry — io.github.stackin-io/stackin, streamable-http, no package to install.

Publishing is automated: pushing a v* tag runs .github/workflows/publish-mcp.yml, which stamps the version from the tag, authenticates with GitHub OIDC (no stored token — the workflow's own identity is what grants the io.github.stackin-io/* namespace) and publishes.

git tag v1.0.0 && git push origin v1.0.0

The registry lists the public repository, which is why the entry lives here and not beside the server's own source.

FAQ

Do I need a digital certificate (A1/A3) in the client? No. The certificate is uploaded once at app.stackin.io and stays server-side.

Does it work in production, or only homologation? Both. The environment is fixed on the API key, or on the connection you authorized over OAuth — there is nothing to pick per call.

Which municipalities are supported for NFS-e? The ones Stackin covers; consult docs.stackin.io. NF-e is nationwide through SEFAZ.

Can I use the server without this plugin? Yes — it is a plain MCP server. The plugin only adds the skills and saves you the manual configuration.

Is it free? Issuing consumes credits from your Stackin account. Consulting and cancelling do not.

The code lookups answered 403 and now they do not. What changed? Until 2026-09-10 those four routes carried no OAuth scope, so an authorized app was refused whatever it had been granted. They now need invoice:read, which every connection already holds — a connection made before that date works without being remade.

A CNPJ I know is valid comes back 404 from lookup_taxpayer. Is it wrong? Probably not. That registry reloads monthly from the Receita Federal's dump, so a company registered in the last few weeks is not in it yet. It is not a validity check, and nothing should block an invoice over it.

Maintainer notes

app_id in .app.json was filled on 2026-09-06, once mcp.stackin.io was registered in OpenAI Developer Mode and a real connector completed the OAuth authorisation. It was left empty until then on purpose — it is the one value that cannot be prepared ahead of time.

A stale tool list was fixed by creating a new app, not by rescanning. The first app (asdk_app_6a9d458d…) had been scanned before the NFS-e restriction and kept advertising the old surface. .app.json now names the second one; if a third is ever created, this file is the thing to update, and nothing warns you when it is wrong.

NF-e came back on 2026-09-07, in two steps. First document_type widened to accept nfe on the existing eight tools — a change the count could not reveal. Then the four NF-e-only operations came back: correct_invoice, invalidate_numbering, list_received_invoices and manifest_received_invoice. Twelve tools then, seven skills.

The four lookup tools landed on 2026-09-10: lookup_fiscal_code, search_fiscal_codes, list_fiscal_kinds and lookup_taxpayer, with the fiscal-lookup and taxpayer-lookup skills. Sixteen tools now, nine skills. They shipped refusing every OAuth token, because their routes carried no scope at all; invoice:read was granted to them on the same day, so nothing has to be reconnected.

The connector caches the tool list from the moment it was scanned. The app was authorized at 10:48 on 2026-09-06 and the NFS-e restriction went live at 12:00, so ChatGPT kept advertising eleven tools — including three the server no longer has. Publishing a new plugin version and running Scan Tools is what refreshes it; the fix is not on the server, which was already correct. Check the count before testing anything: the server is the truth, and scripts/check_skills.py reads it live.

Read the identifier off the connector page, not out of the URL. The address is chatgpt.com/plugins/plugin_asdk_app_…, and plugin_ there is part of the route, not of the id. The page's own "App ID" field is the value that belongs here — asdk_app_…. Getting that wrong is silent: nothing validates it until a submission fails.

mcp.authentication was bearer, written before the server spoke OAuth. The connector page now reports Authorization used: OAuth, so it says oauth. The enum for that field is not in OpenAI's published documentation — developers.openai.com/codex/plugins/apps.md redirects to a 404 — so this is the honest description rather than a verified value. If the submission rejects it, that is where to look first.

License

MIT — see LICENSE.