Turn your company’s files, servers and workflows into context for every AI agent — and into open-weight models you own. Private by default, even with Claude and ChatGPT.
A browser page is reviewed and saved with Velora Sources. In Velora, the team approves it as project knowledge, reviews an agent’s work and its passing checks, and exports reviewed training examples with their sources.
1,000+active users in our first 3 months
$50Kin venture funding committed USD
Company customers
Employee usage
Individual use
Individual use
One app for the intelligence your company owns.
From the first source to a model your agents use, every step is a screen in Velora — on your computer, reviewed by you.
Four steps, all on your computer.
Bring your data, give your agents its context, train a model you own, and decide what hosted models may see — with the next thing to review always in view.
Every source kept with its revisions
Approved knowledge reaches every private task
Privacy chosen per project
intelligence / overview
Your files, pages and servers.
Import documents, capture pages from your browser, or sync your own servers on request. Each source keeps its text, its revisions and who approved it for training.
Server tokens stay encrypted on your computer
Browser capture from Chrome and Edge
Exact source text, never a summary
intelligence / sources
Learn from your day, if you want to.
Off until you turn it on. The app in front and its window title become a daily journal in Sources, and documents in folders you choose are imported when they change. Pause it, or delete everything it captured, in one step.
Never keystrokes or the screen
Password managers and private windows skipped
Nothing trains until you approve it
intelligence / sources
Training examples, drafted for you.
Pick a source and Velora drafts question-and-answer examples from it, each quoting the exact passage it came from. Keep, edit or drop each one, then approve the dataset.
Drafted on your computer, no model needed
Every answer cites its source
Keep personal details for your own model, or take them out
intelligence / datasets
A model you own, with its scorecard.
Fine-tune an open-weight model on your Mac or on another computer you own. Every run is scored against its base model on your own examples, then you can give it to your agents, export it, or train it again.
Held-out examples measured first
Export with a model card
Carry your data to a newer base model
intelligence / models
Train where the model lives.
Pair a computer you own that has the model — a Mac Studio in the office, a Linux box with MLX — and train, score and serve on it from your laptop. Traffic between them is encrypted with a key from its pairing code.
Works on office Wi‑Fi without certificates
Its models appear in every model picker
Local only projects stay on this computer
intelligence / models
Mask detected sensitive values.
Choose Local only, Shielded or Standard for each project. Preview how sample text is masked and see counts by category.
A brief in one chat. A decision in another. Three agents solving different versions of the same problem.
Velora connects the context behind the work, so your team spends less time explaining, untangling overlaps, and putting the pieces back together.
One project brief. One set of decisions.
BuildFollows the approved brief
ResearchUses the current sources
ReviewChecks shared requirements
A shared understanding. Distinct responsibilities.
What your company knows. What your agents need.
A shared understanding starts with the right sources. Build it once, review it, and carry it into the work that follows.
01
Bring in what your company knows.
Files, web pages, workflow decisions — and your own servers, synced on request. Every source stays on your computer, connected to its original text and revision.
Files · Websites · Servers · Workflows
02
Give every agent the same context.
Approve what your agents can use. Requirements, responsibilities and decisions travel with the work as it moves between people and models.
Reviewed knowledge · Clear roles
03
Train a model you own.
Turn reviewed examples into training data, fine-tune an open-weight model on your Mac, and give it to your agents.
Datasets · Open-weight models · Evaluations
04
Keep it private.
Choose how each project uses hosted models. Shielded mode masks detected sensitive values in text.
Before sending text to a hosted model, Velora can replace detected credentials, personal data and your private terms with placeholders. Detection can miss sensitive information, so review content before sending it.
What you write
Email maria@acme.com the Acme Corp renewal, and use key sk-proj-7Hq2… for the billing API.
What the hosted model sees
Email [EMAIL_3fa9c21b] the [TERM_58f7610f] renewal, and use key [SECRET_45414619] for the billing API.
What you get back
Drafted the renewal to maria@acme.com and called the billing API with your key.
Restored on your computer — in the reply, in the files a coding agent changes, and in every tool Velora runs for the model.
Local onlyVelora blocks hosted models for the project; models on your computer can still work on it.
ShieldedThe default. Detected credentials, personal data and private terms in text become placeholders.
StandardHosted models receive text as written, with recognised credentials replaced where detected.
Placeholders come from a key stored on your computer. Coding agents can work in a shielded copy without the project’s Git history, and their changes come back with real values. Binary files pass through unchanged, and agents may access other files if their own sandbox allows it.
Train models you own. Then hand them to your agents.
Open-weight models are now strong enough to specialise. Velora turns the work you already reviewed into a model that knows how your company does things — trained on your Mac, measured on your data, and changed only when you decide.
01Pick an open-weight modelQwen, Llama, Gemma or Mistral, pinned to an exact version, with a license that allows fine-tuning. One click sets up Apple’s MLX.
02Train it on approved dataVelora drafts examples from your sources for you to review, each quoting its source. LoRA fine-tuning then runs on your Mac.
03See if it got betterEvery run is scored against its base model on your own examples — held-out ones first, the ones it never trained on.
04Use it, export it, retrain itServe it to every agent and chat, export a standalone copy with its model card, or carry the same data to a newer base model.
Qwen2.5 0.5B
Gemma 3 1B
Qwen3 1.7B
Llama 3.2 3B
Qwen3 4B
Mistral 7B
Qwen2.5 7B
Training and serving run on Apple silicon Macs, or on another computer you own that runs MLX — including from Windows, by pairing one. Larger models need more memory. A trained model is a candidate — evaluate it before you rely on it.
Train on the computer that has the model, let it learn from the way you actually work, and choose what personal details it may keep — each one a switch, off until you turn it on.
Your laptopOffice Mac Studio
Encrypted with its pairing code
Train on any computer you own
Pair a Mac or Linux box that has the model, and train, score and serve on it from your laptop. Its models appear in every model picker.
Code — server.ts
Linear — Ship beta.17
Notes/standup.md
Password manager — skipped
Learn from your day
If you turn it on, the app in front and its window title become a daily journal in Sources, with documents from folders you choose. Never keystrokes or the screen.
Keep them — the model is yoursTake them out
Detected credentials always come out
Personal details, your call
See how many names, emails and numbers your examples mention, then keep them for a model you own or take them out first. The model card records which.
Hosted models are a great way to start. But a model you rent can change or be retired, and a fine-tune you pay for stays on the provider’s servers. Velora keeps what your company knows — and the models that learn from it — with you.
TopicRenting itOwning it with Velora
Your company’s dataSent as written in prompts to someone else’s model.Train locally, mask detected values in hosted requests, or block hosted models for a project.
What you build upPrompts and chat history.Reviewed knowledge, approved datasets, scorecards and model weights — in folders you own.
When a better model shipsStart again with prompts tuned for the new model.Train the new base model on the same approved data and compare the scorecards.
Who decides what changesThe provider’s release schedule.You. A trained model is a candidate until you choose to use it.
Where it runsThe provider’s cloud.Your Mac, another computer you own, or — exported — anywhere MLX runs.
Company data. Company rules.
You decide which sources become context, what gets shared, and which models do the work.
Hosted models receive the content you send. Shielded mode replaces detected sensitive values in text, but can miss them. Review content before sending it. Local only blocks hosted models in Velora for that project.
Can Claude Code and Codex read my files?
With file shielding on, they work in a copy with detected values masked and no project Git history. Binary files remain unchanged, and an agent may read outside the copy if its own sandbox allows it. Changes come back to your project with conflicts kept separately.
What do I need to train a model?
An Apple silicon Mac, or another computer you own that runs MLX — a Mac Studio in the office or a Linux box — paired with Velora. One click installs Apple’s MLX into an environment only Velora uses. Pick an open-weight model — Qwen, Llama, Gemma or Mistral — and train it on examples you approved. On Windows you can train on a paired computer, or curate and export datasets to train anywhere.
Can I train on a different computer from the one I use?
Yes. Turn on hosting on the computer that has the model, or run velora train-host on a Linux box, and pair with the code it shows. Velora sends the approved examples there, trains, scores and serves on it, and its models appear in every model picker. Traffic between the two is encrypted with a key derived from the pairing code, which is never sent over the network.
Will my own model learn personal details?
Only if you choose. When you create a dataset or start a run, Velora shows how many personal details the examples mention and lets you keep them — the model is yours — or take them out first. Detected credentials are always taken out, and the model card records the choice.
What does Learn from your day record?
Nothing until you turn it on for a project. Then, every 20 seconds, the app in front and its window title go into a daily journal in Sources, and documents in folders you choose are imported when they change. It never records keystrokes or the screen, skips password managers, private windows and anything you exclude, and you can delete everything it captured in one step.
How do I know the trained model is better?
Every finished run is scored against its base model on your own examples, starting with the held-out ones it never trained on. It is a measure on your data, not a general benchmark, so a trained model stays a candidate until you decide to use it.
Can I take the model with me?
Yes. Export fuses the training into a standalone model in a folder you own, with a model card recording its base model, license, dataset and scorecard. It runs anywhere MLX runs, and Velora can serve it to your agents or train it again.
What happens when a better open-weight model comes out?
Your reviewed dataset carries over. Prepare a run on the new base model with the same approved examples, then compare the two scorecards before you switch.
What does it cost?
Velora is free to start. Training, serving and the privacy shield run on your own computer, so a model you own has no per-token bill. Hosted models are billed by the provider you already pay, and Pro and Teams add more parallel work and shared projects.
Make your next agent part of the team.
Start with your knowledge. Keep it private. Own the model that learns from it.
