AI Providers (FiestaBot setup)

FiestaBot is FiestaBoard's built-in AI assistant. Click AI Assistant in the sidebar (the Sparkles icon) to open the chat panel, then describe what you want in plain language — FiestaBot drafts and edits board pages through conversation, and can also install plugins, update settings, and manage schedules.

FiestaBoard ships without any bundled LLM credentials. You bring your own provider, your own API key, and your own model list. There is no FiestaBoard-hosted AI proxy.

Two protocols are supported out of the box:

• OpenAI-compatible chat completions — one-click presets for OpenAI, OpenRouter, DeepSeek, Mistral, Together AI, and Fireworks AI, plus local servers Ollama, LM Studio, llama.cpp, and vLLM. Any other OpenAI-compatible endpoint works too — just paste the base URL.
• Anthropic Messages API — direct access to api.anthropic.com using a Claude API key.

In Settings → AI Providers, Add provider asks what kind of provider you want and then shows only the fields that kind needs.

The preset list is intentionally limited to providers that fully honor the OpenAI response_format: json_object field that the page generator relies on. Other services (Google Gemini's OpenAI shim, xAI Grok, Perplexity Sonar, Cerebras, Nvidia NIM, DeepInfra, …) can still be used by entering the base URL by hand — just be aware that strict JSON mode may need the prompt-only fallback to kick in.

Choosing a provider

Not sure which one to pick? Here's the short version.

We recommend OpenRouter for most users

OpenRouter is our suggested starting point. It's a single API and a single API key that gives you access to hundreds of models from OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek, and others — including most of the open-weight models — with pay-as-you-go billing and no per-provider account juggling. You can swap models from the FiestaBot model dropdown without re-entering credentials.

Disclosure: We recommend OpenRouter because it works well, not because we get paid to. FiestaBoard receives no referral fees, kickbacks, affiliate commissions, or other compensation from OpenRouter — there's no ?ref= link in our docs and no partner agreement of any kind. We just like the product. If you'd prefer a direct relationship with one provider, any of the other presets works equally well.

When to pick something else

• You already pay for OpenAI / Anthropic / Mistral / etc. — use the matching preset and your existing key. No reason to add an OpenRouter middleman.
• You want the lowest possible latency — DeepSeek is typically faster than OpenRouter's pooled routing.
• You want everything to stay on your own hardware — pick a local preset (Ollama, LM Studio, llama.cpp, vLLM). Note the caveats about smaller models and JSON adherence below.
• You're in the EU and care about data residency — Mistral is EU-hosted.

Quick setup

In Settings → AI Providers, click Add provider and choose what kind of provider it is:

• Sign in: ChatGPT, OpenRouter, or Hugging Face. The provider is saved straight away and shows its sign-in panel. Once you are signed in, its models are listed on their own: search the list and pick one. See Signing in instead of using an API key.
• Use an API key: pick the service (OpenAI, Anthropic, DeepSeek, Mistral, Together AI, Fireworks AI, or OpenRouter), then paste your key and add a model. The name, base URL, and protocol are filled in for you.
• Run it on my network: pick Ollama, LM Studio, llama.cpp, or vLLM. The Server address is filled in with that server's usual port on localhost, which means the board itself. If the server runs on another computer, change it to that computer's address, such as http://192.168.1.20:11434/v1. Then add a model.
• Advanced (custom endpoint): the full form, for any other OpenAI-compatible or Anthropic endpoint. Its Quick presets pills fill in the name, base URL, and protocol for a known service.

Opening a provider later shows the same short view it was set up with. Everything else (name, protocol, base URL, API key, sign-in) is under Advanced in that view. A provider that matches none of the kinds opens in the full form.

Configuration

1. Open Settings → AI Providers.
2. Toggle the top switch to Enabled.
3. Click Add provider, choose a kind (see Quick setup), and fill in what it asks for. The full form (Advanced (custom endpoint), or Advanced inside any provider) has every field:
   • Name: any label, e.g. OpenRouter or Claude. Auto-filled when you pick a service or click a preset pill.
   • Protocol: pick OpenAI-compatible or Anthropic. The quick-pick buttons below also set this for you.
   • Base URL: the API root, e.g. https://openrouter.ai/api/v1 or https://api.anthropic.com/v1.
   • API Key: paste the key. It is stored on this device's data/config.json and is masked (***) on read.
   • Models: search the provider's list and pick, or type an id the list does not have (e.g. openai/gpt-5-mini, claude-sonnet-5) and choose Use this id. A signed-in provider lists its models as soon as you are signed in; any other saved provider lists them when you press Load models. Refresh models asks again.
   • Default model: picked automatically once you add at least one model; change it from the same searchable list.
4. (Optional) Click Test connection to send a one-token smoke test and confirm credentials and connectivity.
5. Click Save changes.

If you configure more than one provider, mark one as the default using the Make default button. FiestaBot will use the default provider unless you pick another from the dropdown in the chat panel.

Signing in instead of using an API key

Three providers also let you sign in with your account instead of pasting a key. This is an extra choice, not a replacement: a provider set up with an API key keeps working exactly as before, and you can switch back at any time.

Sign in with	Where requests go	What you need
OpenRouter	https://openrouter.ai/api/v1	An OpenRouter account. Signing in creates an API key for FiestaBoard in your account.
Hugging Face	https://router.huggingface.co/v1	A Hugging Face account. FiestaBoard asks only for permission to call inference.
ChatGPT	https://api.openai.com/v1 (OpenAI Responses)	A ChatGPT account. Requests use the OpenAI Responses protocol, which is set for you.

To sign in:

1. Open Settings → AI Providers, click Add provider, choose Sign in, and pick ChatGPT, OpenRouter, or Hugging Face. The provider is saved with the right base URL and protocol, and shows a sign-in panel.
2. Press the sign-in button in that panel and approve FiestaBoard on the provider's page. OpenRouter and Hugging Face bring you back to Settings signed in; ChatGPT is finished by a paste (below).
3. Pick the model to use from the list that appears (search it by name or id). FiestaBot can use the provider at once, and its own model picker offers the same list.

An existing API-key provider can switch to a sign-in too: open it, then under Advanced (or in the full form) choose a service under Or sign in instead of using an API key and press Save changes. Your API key field is left as it is.

ChatGPT cannot send you straight back to your board, so the sign-in is planned around that:

1. Press Sign in with ChatGPT. Settings shows a Finish signing in step with these instructions and a box to paste into. Read them, then choose Open the ChatGPT sign-in page. ChatGPT opens in a new tab.
2. Sign in and approve FiestaBoard in the ChatGPT tab. That tab then ends on an error page, such as "can't connect to 127.0.0.1" or "This site can't be reached". That is expected: nothing went wrong.
3. Copy that tab's whole address from the address bar (it starts with http://127.0.0.1), come back to the FiestaBoard tab, paste it into the Finish signing in step, and press Finish sign-in. The panel then says you are signed in.

This paste step goes away once OpenAI approves FiestaBoard's app: ChatGPT then sends you straight back to your board, like the other providers.

For the other providers, a sign-in that does not come back by itself can be finished the same way: choose Sign-in didn't come back? Paste the address or code.

OpenRouter can also be signed in without a browser redirect: choose Sign in without a browser redirect, approve, and paste the code OpenRouter shows.

While a provider signs in, its API key is not used. Press Use the API key instead (under Advanced) to go back to the key; nothing you typed there was removed. Removing a provider, or switching it back to a key, deletes the sign-in FiestaBoard was holding for it.

Sign-ins are stored like plugin sign-ins: in data/oauth_tokens.json on your board, never in data/config.json, never sent to your browser, and not part of backups. After restoring a backup onto a new board, sign in again. If a provider stops accepting the sign-in, FiestaBot says "Sign in to provider again in Settings → AI." and sends nothing until you do. See Connecting Accounts for how the sign-in itself works.

Recommended models

These all work well with the FiestaBoard prompt format. Where a provider offers a stable alias (-latest, or an undated family name like claude-sonnet-5), we prefer it over a date-stamped id so the recommendation ages more gracefully.

Provider	Protocol	Model	Notes
OpenRouter	OpenAI	openai/gpt-5-mini	Cheap, fast, reliable JSON output.
OpenRouter	OpenAI	anthropic/claude-sonnet-5	High-quality, slower.
OpenAI	OpenAI	gpt-5-mini	Same as via OpenRouter.
Anthropic	Anthropic	claude-sonnet-5	Direct, no OpenRouter markup.
Anthropic	Anthropic	claude-haiku-4-5	Cheaper, fast.
DeepSeek	OpenAI	deepseek-chat	Cheap, capable.
Mistral	OpenAI	mistral-large-latest	EU-hosted option.
Together AI	OpenAI	meta-llama/Llama-3.3-70B-Instruct-Turbo	Hosted open weights.
Fireworks AI	OpenAI	accounts/fireworks/models/llama-v3p3-70b-instruct	Hosted open weights.
Local Ollama	OpenAI	qwen2.5:14b-instruct or larger	Needs a model that follows JSON.
Local LM Studio	OpenAI	A 14B+ instruction-tuned model	Same JSON-adherence caveat.

Note: Model catalogs change often, and providers retire ids on their own schedule. Treat this table as a starting point — check the provider's current model list, and if Test connection reports an unknown model, pick the closest current id from their catalog.

Smaller (≤ 7B) local models often struggle to emit valid JSON for the FiestaBoard schema; if you see frequent "Could not parse JSON" errors, switch to a larger or instruction-tuned model.

Privacy

Each time you send a message, FiestaBot sends to the provider you configured:

• The system prompt (board dimensions, character set rules, JSON schema, and available template functions).
• Your message.
• The variable list of all enabled plugins (names + descriptions
  • max widths). This may include data such as transit station IDs or location names that you have configured.
• Up to a handful of example pages drawn from plugin manifests.
• The current page being edited (when a page is open in the editor — FiestaBot always has context of what you're working on).
• Context about your pages, schedules, carousels, and installed plugins (names and IDs only, not content).

API keys are stored locally in data/config.json and never sent to any FiestaBoard-hosted service.

Chat history

Every FiestaBot conversation is saved on your FiestaBoard as you go — there is nothing to click. Closing the drawer or reloading the page brings you back to the chat you were in, and the History button in the panel header (the clock icon) lists everything you have talked about, newest first, with a search box for the title.

• Review — open any conversation to read it back, including the actions FiestaBot took and what each one returned. Nothing runs while you are reading.
• Continue — picks a saved conversation up again as the live chat. Its provider, model and "don't ask again" choice come back with it.
• New chat (the pencil icon) starts a fresh conversation; the old one stays in History.
• Rename and Delete sit on each row. Clear all at the bottom removes every saved chat after a confirmation.
• Export JSON on an open conversation downloads that thread as a file, handy for attaching to a bug report.

The title is the first thing you typed, trimmed to 80 characters; rename it whenever you like. FiestaBoard keeps the 200 most recently used conversations and drops the oldest when a new one arrives. Credentials never end up in a saved chat: any API key, password or token that passes through a tool is stored as ***, the same masking the tools themselves apply. History lives in data/ai_conversations.json and is not part of the settings backup.

Limitations

• Two protocols supported: OpenAI-compatible chat completions (presets cover OpenAI, OpenRouter, DeepSeek, Mistral, Together, Fireworks, Ollama, LM Studio, llama.cpp, and vLLM), and the Anthropic Messages API. Other providers (Google Gemini, xAI Grok, Perplexity, Cerebras, DeepInfra, Nvidia NIM, Cohere, …) can be reached today through OpenRouter, by entering the base URL manually (some don't strictly honor JSON mode), or by registering a new entry in src/ai/protocols.py.
• No image/vision input.
• FiestaBot acts through the same tools as the MCP server and shows each action as it happens. Creating, editing and configuring apply immediately. Destructive actions — every tool the MCP server marks destructive except the system tier: deleting a page, schedule, collection, plugin instance, board or panel, uninstalling a plugin, forgetting or disconnecting Wi-Fi — follow the chat's approval mode, the Ask / Auto toggle in the panel header. In Ask (the default) each one pauses for your Approve / Deny; in Auto they run without asking and are marked auto in the step list. The approval card's "Approve and don't ask again in this chat" is the same thing for one conversation, until you start a new chat. Restarting, shutting down and updating the system always ask, in every mode, and FiestaBot cannot change the approval mode itself. Stop ends a turn at any time.
• A modest per-process rate limit applies to page generation requests to protect against runaway clients (1 second between calls, 2 concurrent).

Troubleshooting

AI Assistant doesn't appear in the sidebar You need at least one provider configured in Settings → AI Providers with the top toggle set to Enabled. The sidebar item only appears once a provider exists.

"Could not parse JSON" or repeated empty results The model isn't returning valid JSON for the board schema. Switch to a stronger model (e.g. gpt-5-mini or claude-sonnet-5) or one explicitly tuned for instruction following.

Test connection fails Double-check the Base URL (it should usually end in /v1), verify the API key has not been revoked, and confirm the model id exists for that provider.