All documentation

Proxyma Desktop

Semantic search and AI chat over your own documents, running entirely on your own machine. No account, no server, no cloud.

Running Proxyma for a team?

Proxyma Server is the multi-user edition, with accounts, roles and shared sources. It has its own manual - Proxyma Server.

What Proxyma Desktop is

You point it at folders and systems you already use, it indexes them, and you ask questions in plain language. Answers cite the documents they came from.

Everything happens on your computer. There is no sign-in, your documents are never uploaded, and with a local AI model configured, nothing you type or index leaves the machine at all.

Installing

Windows

  1. Download proxyma-x64-v<version>.exe from the releases page.
  2. Run it. The installer bundles its own Java runtime and Python tooling, so there is nothing else to install.
  3. Launch Proxyma from the Start menu. It opens in your browser.
Windows SmartScreen

The installer is not yet signed with an Extended Validation certificate, so SmartScreen may warn on first run. Choose More info, then Run anyway. Verify the download came from the official releases page first.

Linux and macOS

Not yet released. The application itself is cross-platform - the installers are what is missing.

What it needs

ComponentRequirement
Operating systemWindows 10 or 11, 64-bit
Memory8 GB, 16 GB if you run a local AI model on the same machine
Disk3 GB for the application, plus room for your index
InternetOnly for licence activation, updates, and a hosted AI provider if you choose one

First run

Proxyma opens on the chat page with nothing indexed yet. Two things to do:

  1. Choose an AI provider - Settings → Providers. Nothing works until one is reachable.
  2. Add a folder - Resources → Add Resource → Files. Start small so you can watch it work.

Then ask a question. If the answer cites your document, you are finished setting up.

Choosing an AI provider

Proxyma needs two kinds of model, and they need not come from the same place:

ModelWhat it does
EmbeddingTurns document text into vectors so search can find things by meaning
ChatReads the retrieved passages and writes the answer

Fully local

Install Ollama and point Proxyma at http://localhost:11434. Nothing you index or ask ever leaves your machine. This is the right choice for confidential material.

Expect it to be slower, and expect answer quality below the hosted models - a model that fits on a laptop is a smaller model.

Hosted

Paste an API key for Anthropic, OpenAI, DeepSeek, Mistral, Azure OpenAI or Google. Better answers, and in exchange the passages relevant to each question are sent to that provider.

Worth being clear about

With a hosted provider, document excerpts leave your machine. Your API key is stored encrypted using Windows DPAPI, so it is protected at rest - but that protects the key, not the content of your questions. If the documents are confidential, use a local model.

Changing the embedding model means reindexing

Vectors from different embedding models cannot be compared. If you switch, everything already indexed must be indexed again before search works properly.

Adding documents

Resources → Add Resource.

SourceWhat you provide
FilesA folder. Proxyma indexes it in place - your files are not copied or moved.
JiraBase URL and an API token
ConfluenceBase URL, space key, API token
BitbucketBase URL, repository, access token
SharePoint / TeamsA registered application with read permission
SVNRepository URL and credentials

PDFs, Office documents, Markdown, plain text and HTML are all handled. Scanned PDFs go through optical character recognition automatically. Tables in PDFs are extracted as tables rather than flattened into a line of numbers.

Credentials are encrypted with Windows DPAPI, tied to your Windows account.

Watching it work

Each document moves through detection, download, conversion and embedding. Large folders take a while - conversion and embedding are the slow parts, and a local embedding model is slower than a hosted one. One document failing never stops the rest.

Re-syncing only processes what changed, so the second run over the same folder is fast.

Using Proxyma

Asking questions

Type a question the way you would ask a colleague. Proxyma finds relevant passages first, then answers from them, and shows you which documents it used.

If nothing relevant is indexed it says so rather than inventing an answer. That is deliberate.

Conversations

Chats are kept so you can return to them. They live in a local database on your machine and are never uploaded.

Search

Keyword matching and meaning-based matching run together, so an exact error code and a vague description of the same problem both find the right document.

Connecting other tools

Proxyma speaks the Model Context Protocol, so editors and AI assistants that support MCP can search your indexed documents directly. The Docs page inside Proxyma has a ready-made configuration to paste into your editor.

Your licence

Proxyma Desktop has a free Community tier and paid Personal and Professional tiers. Each is a one-time purchase covering one installation - there is no subscription and no renewal.

 CommunityPersonalProfessional
PriceFree$29 once$49 once
Local folders1UnlimitedUnlimited
Files per folder100UnlimitedUnlimited
Index subfoldersNoYesYes
Websites1UnlimitedUnlimited
Crawl depth / pages per site3 / 10050 / 2000050 / 20000
Search, chat, any AI modelYesYesYes
Jira, Confluence, Bitbucket, SVNNoNoYes
SharePoint, OneDrive, TeamsNoNoYes

User accounts, roles and single sign-on are not part of any desktop tier - they belong to the Enterprise server edition, which is a shared deployment rather than a per-computer licence. An Enterprise licence is refused on the desktop edition, because there is nothing there for it to unlock.

A licence does not include AI usage

Proxyma needs an AI model to answer questions, and you provide it. Running one locally with Ollama costs nothing beyond your own hardware and keeps everything offline. Connecting a cloud provider instead - Anthropic, OpenAI, DeepSeek, Mistral, Azure OpenAI or Google - uses your own account and API key, and that provider bills you directly at its own rates. No part of your licence fee goes towards it, and we never sit between you and their bill.

Where your licence key comes from

Buy through the pricing page on proxyma.ai. Payment is handled by Lemon Squeezy, who act as the merchant, charge any applicable tax and send your receipt. Your licence key is in that receipt email, and it stays available on their My Orders page if you lose it.

Activating

After purchase you receive a licence key by email. Enter it in Settings → Licence.

Activation needs an internet connection - once. Proxyma contacts the licence server, registers this installation, receives your licence and stores it on your machine. That is the only step that waits for the network.

After that, Proxyma re-checks the licence quietly in the background every few hours. This is how a refunded or revoked licence stops working. It never blocks you. If the check cannot be made - you are offline, behind a firewall, or our server is down - nothing changes for 14 days. In the last few of those days a banner appears asking you to get online. If the 14 days run out, Proxyma keeps running with the free Community tier's features until it can check again; nothing you have indexed is deleted and no settings are lost. The moment a check succeeds, everything you paid for comes straight back, with nothing to re-enter.

Why Proxyma checks at all

Each licence covers one installation, and that can only be counted somewhere that sees all of them. The recurring check is what lets a refund actually take effect - without it, asking for your money back would leave you with a working licence forever, which is not a product we could sell honestly. It is built to fail in your favour: an outage on our side delays a revocation, it never stops a paying customer working.

Only your licence key and a random identifier for this installation are sent - plus your computer's name, at activation only, so you can tell your machines apart. Never document content, never anything about your hardware, never usage data, never what you have searched for.

The free Community tier needs no licence and makes none of these calls.

Moving to another computer

Settings → Licence → Deactivate. This releases the installation so you can activate the same key elsewhere, and returns this copy to the Community tier.

If you cannot reach the deactivate button - the machine died, or was wiped - the seat can be released for you. Contact support with your licence key.

Reinstalling counts as a new installation

Proxyma identifies an installation by a random value created when it is first set up, not by your hardware. That is deliberate: hardware identifiers change when you swap a dock or clone a disk, and people were being locked out of computers they had never replaced.

The consequence is that reinstalling on the same computer looks like a second device. Deactivate before you reinstall, or release the old one afterwards.

Updates

Proxyma checks for updates and tells you when one is available. Updates preserve your index, conversations and settings.

Where your data lives

Everything is under your user profile:

WhatWhere
Index, conversations, settings%LOCALAPPDATA%\\Proxyma\\
CredentialsWindows DPAPI, tied to your Windows account
Your documentsWherever they already are - Proxyma does not copy them

Back up the Proxyma folder to keep your index and chat history. Losing it costs you a reindex, not your documents.

Troubleshooting

It will not start

Proxyma bundles a vector database that starts alongside it. If startup stalls, check that no antivirus or endpoint protection is blocking the bundled executables in the installation directory.

"No AI provider configured"

Settings → Providers. For Ollama, confirm it is running: open http://localhost:11434 in a browser and expect a response.

Indexing is very slow

Embedding is the bottleneck, and a local model on a laptop CPU is slow. Options: use a hosted embedding provider, index fewer folders, or leave it running overnight - it resumes where it left off.

Search does not find something I know is there

Check the document actually finished indexing - Resources shows per-document status. Scanned documents depend on OCR quality. If you changed the embedding model, reindex.

Answers are wrong or vague

Proxyma can only answer from what it has indexed. If the source material is thin, so is the answer. A larger chat model usually helps more than any setting.

My licence key is rejected

Proxyma tells you which of three different things went wrong, because the fix differs:

What it saysWhat to do
Already active on 1 of 1 devices The key is fine and in use elsewhere. Deactivate on the other machine, or ask support to release it.
Could not reach the licence server The key is probably fine - this is a network problem. Check your connection and any corporate proxy, then try again.
Not recognised, revoked or expired Check for whitespace when pasting. If it still fails, contact support.

Proxyma stopped working after I went offline

Search, chat and your indexed documents all keep working offline indefinitely - if one of those stopped, it is a connector or a cloud AI provider that needs the network, not your licence.

Paid features are a separate matter. Proxyma re-checks your licence in the background, and after 14 days with no successful check it falls back to the Community tier's limits. You will have seen a banner for several days before that happened. Connect to the internet and press Try again in that banner, or reopen Proxyma; everything comes back immediately and nothing was lost.

The banner says my licence was revoked

That normally means the order was refunded or charged back. If you did not ask for a refund, email [email protected] with your licence key and we will sort it out.

Uninstalling

Use Add or remove programs. This leaves your index and conversations behind; delete %LOCALAPPDATA%\\Proxyma\\ to remove those too. Your own documents are untouched either way.