Many minds, one refined result.
WaxFrame puts multiple AI assistants to work on your document at the same time. One acts as the Builder — writing and refining. The rest act as Reviewers — giving specific, numbered suggestions. Round by round, your document gets better. Whether you're starting from scratch, refining an existing document, or generating the text for a presentation or any other format — the hive handles it.
Each AI in your hive acts as a reviewer — it reads your document and suggests improvements each round.
Set up at least two AIs to continue. Three or more is recommended — with three or more reviewers the hive can resolve disagreements on its own; with only two, you'll occasionally have to break a tie yourself.
Click any AI card below to set it up. New here, or wondering what it costs to run? Tap the ⓘ next to the heading above — cost, keys, and the full setup guide all live there.
The Builder is the AI that actually rewrites your document each round. While your reviewers give numbered suggestions, the Builder reads all of those suggestions plus the full document, then produces the updated version. Because it handles the entire document every single round, the Builder uses significantly more tokens than reviewers — it must have a paid API subscription with enough capacity. Choose an AI you trust with large documents. You can change your Builder any time from the work screen.
Think of the Builder as a handyman as well as a synthesizer. Between full rounds, you can write a quick directive in the Notes drawer — "rewrite the first paragraph," "stop using the word 'refine' so much," "tighten paragraph three" — and click Send to Builder to have it perform just that one task without running the full hive. Saves tokens and time. Then click Smoke the Hive when you want the reviewers to weigh in on the change.
Claude, ChatGPT, Gemini, and DeepSeek all handle large documents well — DeepSeek in particular is extremely cost-effective as a Builder. Avoid free-tier AIs as Builder — they run out of tokens fast. If a round fails with a missing output structure error, try switching to ChatGPT or Gemini. Learn about tokens and how to save money →
This is the most important page in the whole setup. This section defines what your project is all about. Every AI sees this information on every round — it's their entire understanding of what you're trying to build. The more specific you are the better your project will be (garbage in, garbage out). A few extra minutes here pays off in every round that follows.
New to WaxFrame? Click 📋 Use Template above, pick Starting from scratch, then choose ⭐ Quick Start — a small, low-stakes project where the hive writes and refines a real chocolate-chip-cookie recipe across a few rounds. Same complete flow as a job description or business proposal; the topic is just harmless enough to focus on learning the workflow. Once done, come back and pick the template that matches your real document.
Fields marked * are required to continue.
What kind of document is this? Be specific — cover letter not document. This is the first thing every AI reads and it sets the format, length, and tone for the whole session. If there is one thing the AIs should never forget, put it here.
Who will read this? The AIs use this to calibrate tone, vocabulary, and how much to explain. IT Director and VP of Facilities gives very different results than general public. The more specific you are, the more consistent every round will be.
What do you want to happen after someone reads this? Approve the budget, schedule an interview, understand the three options and pick one. This is the finish line the AIs are working toward. Vague here means vague everywhere — this field has more impact on quality than any other.
Draw the fence. AIs expand into adjacent topics unless you explicitly stop them. Cover both sides — what must be included and what must not be included. Example: wireless networking only — no switching, routing, or security products.
How should it sound? A few adjectives — professional, confident — is a floor that works but leaves a lot open to interpretation. Richer is better: Direct and confident, not stiff — like a peer they'd want to work with. Say what to sound like, what to avoid, and who's reading. Bland inputs produce inconsistent outputs across rounds.
Hard rules that must survive every round — things the AIs should always or never do. Example: do not add new sections, always use Oxford commas, never change the numbers in the cost table. You can also use the Notes drawer on the Work screen for one-round-only instructions.
Pick a mode that matches what your document needs. The Builder is told the rule for the mode you choose, and round-end and convergence-time guards fire accordingly.
Optional. Source material The Hive will reference on every round but never edit — job descriptions (rules and responsibilities), RFP requirements, style guides, scoring rubrics, prior decisions, vendor claims. Distinct from Notes (round-to-round directives) and from your Starting Document (the document under construction).
Three buttons below let you choose where your starting document comes from. Click Upload File to upload a Word doc, PDF, or other file you already have. Click Paste Text to paste text directly. Click Start from Scratch if you want the hive to generate a first draft using only your project goal.
Click the area below to browse for a file, or drag and drop one directly onto it.
Your preferences are saved in this browser and apply to every project on this machine. Changes save automatically.
How the hive behaves during hands-off Auto runs. These only take effect while Auto is on.
Which AI reads text out of scanned or garbled PDFs — used by the "Re-extract with AI Vision" button and the automatic pass on image-only pages. It uses that provider's configured model.
More sections (autosave frequency, and more) are on the way. The Auto behavior that consumes these options ships alongside the "Auto really means Auto" work.
The full address of the provider's chat-completions endpoint. Most OpenAI-compatible providers end with /v1/chat/completions. Examples:
https://api.mistral.ai/v1/chat/completionshttps://api.together.xyz/v1/chat/completionshttp://localhost:11434/v1/chat/completions (Ollama running locally)Anthropic and Google use different URL shapes — Quick Add handles those automatically.
How the provider expects requests. Most providers (Mistral, DeepSeek, Together, Cohere, Ollama, LM Studio, etc.) speak OpenAI compatible. Anthropic and Google have their own formats — pick those only if you're connecting directly to them.
The secret token the provider gave you when you signed up. Usually starts with a prefix like sk-. Local providers (Ollama, LM Studio) don't need a key — leave blank. Your key is stored in your browser only and never sent anywhere except the provider's own endpoint.
The specific model on that provider's API. Click Fetch Models after entering the URL and key — WaxFrame will pull the live list and ask the provider which model is best as a Reviewer (✨) and which is best as a Builder (🔨). You can pick any model from the list.
Most providers WaxFrame doesn't list still work — they just need to be filled in manually. If the provider's docs say "OpenAI-compatible," set API Format to OpenAI compatible and use whatever URL their docs give you. If it's not OpenAI compatible, you'll need a provider that is, or wait for WaxFrame to add native support.
Import from Model Server works with any AI platform that exposes an OpenAI-compatible chat completions API. Fill in the three fields on the left and click Fetch Models.
/api/chat/completions. For Ollama and LM Studio it ends in /v1/chat/completions./api/models. For Ollama it is /api/tags. For LM Studio it is /v1/models.
Pick any image file (PNG, JPG, SVG, WEBP). It'll be resized to 256×256 max and stored inline with the AI — no external URL, safe for air-gapped networks.
Export your document before starting a new project — your working document and round history will be cleared when you do. If you want to review which reviewers were still suggesting changes (the holdouts), make a manual edit, or run another round before exporting, click ← Back to Hive at the bottom.
Your WaxFrame Pro license is active.
Change BuilderThe Builder rewrites the document each round. Choose an AI with a paid API key and enough token capacity. Cost note: the Builder reads project setup, reference material, the full working document, and every reviewer's suggestions every round — switching to a premium model here can raise per-round cost significantly. If you've enabled billing on Google AI Studio, Gemini-as-Builder bills per token.
This is the most important section of the entire setup. Every AI in your hive reads these six fields before touching your document — it is the only context they have about what you actually want. Vague fields produce a vague document. Specific, detailed fields produce something you can use.
You do not need to fill all six — even two or three give the hive a strong foundation. Fill them top to bottom. Each one controls a specific dimension of the output:
The goal works the same way whether you are uploading a file, pasting text, or starting from scratch. The AIs have nothing else to work from — no internet access, no knowledge of your organisation, no understanding of what you want unless you put it here.
Read the full guide in the user manual — Step 3: Define Your Project ↗
Templates fill in your Project fields with proven starting content for common document types. Pick one matching what you're writing, review the filled-in fields, and edit anything that doesn't fit your specific situation.
The Test All Keys check sends a one-word prompt (Reply with exactly one word: CONNECTED) to every saved AI and reports the result. A red ✗ means something blocked the request. Use the table below to identify the cause from the row's status code, then jump to the matching provider billing console at the bottom of this guide.
Access-Control-Allow-Origin header that permits weirdave.github.io. WaxFrame runs entirely in your browser and cannot bypass CORS. You'll see this almost exclusively with custom endpoints — internal company gateways like Alfredo, self-hosted Open WebUI / Ollama, or other proxies that haven't whitelisted browser origins. Fix is on the gateway side: have the operator add Access-Control-Allow-Origin: * (or specifically https://weirdave.github.io) to the gateway's CORS configuration. The 6 default providers all have CORS pre-configured.
Provider billing & console links — open the right one for whichever AI is failing.
Custom AIs added via Add Custom AI or Import from Model Server won't appear above — only the 6 default providers have known billing URLs. For custom endpoints, billing/permissions are handled wherever you got the endpoint from.
—
—
—
—
Every WaxFrame session has two AI roles and one document, processed in rounds. Each round is the same loop: reviewers read, Builder rewrites, you decide what to keep.
Quoting exact contested text in Notes is more reliable than citing line numbers — line numbers shift as the document evolves.
The Menu is the navigation drawer you'll find at the top-left of every screen. The top section jumps you between setup screens and the Working Console; below that, four grouped sections hold tools, documentation, and advanced features. None of it is required to run a round, but most of it earns its place once you've used the app for a while.
Navigation
Tools
Documentation
Advanced
Documentation pages open in new browser tabs — your in-progress WaxFrame session won't be disturbed. The Menu also has an ℹ️ About button at the top with credits, build info, and license details.
Your Worker Bees are the AI assistants that act as reviewers each round. They each read your document independently and submit a numbered list of specific suggestions for improvement. The Builder then reads all of those suggestions and rewrites the document.
weirdave.github.io is allowed to read responses from a given API endpoint. The 6 default providers (Anthropic, OpenAI, Google, DeepSeek, xAI, Perplexity) all whitelist browser clients, so you'll never see CORS errors with default keys. You'll see them with custom endpoints — Open WebUI / Ollama / internal gateways like Alfredo — when the gateway operator hasn't added Access-Control-Allow-Origin for browser callers. The fix is on the gateway side, not on WaxFrame's side. Click the ⓘ on Test All Keys for the full troubleshooting reference.
Don't have keys yet? Open the API Key Guide ↗ for step-by-step instructions and direct billing links for each provider.
Read the full guide in the user manual — Step 1: Set Up Your Worker Bees ↗
The Builder is the AI that actually rewrites your document each round. While your reviewers each read the document and give numbered suggestions, the Builder reads all of those suggestions plus the full document and produces the updated version.
See What Are Tokens? ↗ for a full breakdown of token costs across providers and how to keep your usage costs low.
Read the full guide in the user manual — Step 2: Choose Your Builder ↗
This is the live version of your document. Every round, the Builder rewrites it in full based on reviewer suggestions. You can edit it directly at any time between rounds — the AIs reference line numbers, so the numbering updates automatically as you make changes.
Read the full guide in the user manual — Step 7: The Work Screen ↗
Pick a mode that matches what your document needs. The Builder is told the rule for the mode you choose, and round-end and convergence-time guards fire accordingly.
Length is measured in your chosen unit — characters, words, paragraphs, or pages. Pages cannot be measured directly from text, so they're approximated using 600 words per single-spaced 12pt page; the other three are direct-counted.
Read the full guide in the user manual — Step 3: Define Your Project ↗
When AIs disagree on something, the Builder flags it here rather than making an arbitrary call.
The panel always shows conflicts from the most recent round only — it is not a project-wide completion indicator. "No conflicts" can mean any of three things: the round was a Builder-Only round (no reviewers ran), the round was Round 1 of a from-scratch project (the Builder writes the first draft before reviewers have anything to evaluate), or the round was a full reviewer round and everyone happened to agree on the changes. None of those mean the document is finished.
The document is "done" when the hive reaches ✓ Converged — a majority of reviewers agree there are no more changes needed. Before that point, keep running rounds with Smoke the Hive, or click 🏁 Finish if you're already satisfied with the current draft.
Click the Current: label on any conflict card to jump directly to that line in the working document.
Read the full guide in the user manual — Step 9: Review Results and Resolve Conflicts ↗
The Live Console shows everything happening in real time — which AIs are sending, responding, succeeding, or failing, with timestamps and response previews.
Use the Copy button to grab the full console log for troubleshooting.
Read the full guide in the user manual — Step 7: The Work Screen ↗
Your hive is the team of AIs collaborating on your document each round. Every checked AI with a saved API key participates automatically. AIs without a key are skipped silently.
Read the full guide in the user manual — Step 7: The Work Screen ↗
The filename mask lets you customise the name of your exported .txt file using tokens replaced automatically at export time.
{name}_final_{version} with project name "My Report" and version "v2" produces My_Report_final_v2.txt. Leave the mask blank to use the default format: ProjectName_version.txt
Read the full guide in the user manual — Step 11: Export and Finish ↗
The three buttons — Upload File, Paste Text, and Start from Scratch — are mode selectors, not action buttons. Click one to choose how you want to provide your starting document.
PDF tip: PDF extraction works best with digitally-created documents. Heavily designed or scanned PDFs may extract poorly — try Paste Text if that happens.
Read the full guide in the user manual — Step 5: Provide Your Starting Document ↗
Notes are directions for the Builder only — not sent to reviewers. They apply to the next round only and are cleared automatically after that round runs. Use Notes for one-round-only instructions; for rules that must survive every round, put them in Additional instructions on the Project screen instead.
Notes pair powerfully with the Send to Builder button: write a directive ("rewrite the first paragraph," "stop saying 'refine'", "tighten the conclusion"), click Send to Builder, and the Builder performs just that task without running the full hive — like asking a handyman for a specific repair. After it's done, click Smoke the Hive to see if the reviewers agree with the change.
Quoting the exact text from the document is more reliable than citing line numbers — line numbers shift as the document evolves, but exact text is stable.
Read the full guide in the user manual — Step 10: Run More Rounds ↗
Reference material is source documentation the hive consults every round but never edits — job descriptions (rules and responsibilities), RFP requirements, style guides, scoring rubrics, prior decisions, vendor claims. It is sent to every reviewer and to the Builder so they can cite against it.
Trim reference material to what is most important. The hive cannot read what you do not paste — but it also costs you money for every token.
Read the full guide in the user manual — Step 4: Reference Material ↗
The token count shown is an estimate based on characters / 4, which is the standard rule of thumb for English text in OpenAI-family tokenizers. Real tokenizers vary by model, language, and content type.
Use this number to gauge cost order of magnitude — a 5,000-character reference doc is roughly 1,250 tokens per request, sent to every reviewer every round. A 6-bee hive running 4 rounds = 30,000 tokens just for reference material, before any document or response content.
WaxFrame is built for desktop work. Minimum viewport is 1366 × 768 px; 1600+ wide is recommended for the multi-panel work screen.
Please resize your browser window, dock dev tools to the bottom, or open WaxFrame on a larger desktop or laptop display.
About
Built with ❤️ by WeirDave and Claude.
