Find the answer across your files, Jira and Confluence, then have the agent act on it.
I have started two sub-agents in the background: one is researching what we decided about the freeze, the other is reviewing pull request #482. I will pick this up when they report back.
Yes, once one fix is in. The freeze starts on August 25, not the 18th: it moved at the release meeting on July 14 (Meeting notes 2026-07-14). Pull request #482 is sound apart from its retry loop, which has no upper bound. I have commented on REL-1182 and corrected the date on the Release Process 2026 page.
Type @ to mention a source or / to use a skill
AI can make mistakes. Always verify important information.
- 1.Every new installation of Desktop or Home Server includes 30 days of Professional, free, once per computer.
- 2.No sign-up: pick an account on the demo's sign-in page. A real Enterprise Server, wiped every night, with a fictional client's key account ready for review.
- Files
- Web pages
- Network shares
- Confluence
- Jira
- Bitbucket
- SVN
- SharePoint
- Microsoft Teams
- Outlook
- Files
- Web pages
- Network shares
- Confluence
- Jira
- Bitbucket
- SVN
- SharePoint
- Microsoft Teams
- Outlook
What did we decide about the Q3 release freeze?
Somebody asks. The answer exists, split between a ticket, a wiki page, a spreadsheet and a chat thread.
text ~ "release freeze"14 issues- REL-1182Freeze exception: payment timeout hotfixIn review
- REL-1175Confirm the Q3 freeze datesDone
- REL-1160Cut the 4.2 release branchOpen
release freeze23 results- Release Process 2026Engineering, updated 3 months ago
- Meeting notes 2026-07-14Release train
- Q3 freeze (draft)Personal space
Q312 files- Q3-plan-v4-final.docxModified 2026-07-12
- Q3-plan-v4-final-FINAL.docxModified 2026-07-15
- release-calendar-2026.xlsxModified 2026-08-02
freeze38 messages- #release-trainAre we still freezing on the 18th?July 11
- #release-trainMoved to the 25th, see the notes.July 14
- HanaIs REL-1182 allowed in during the freeze?August 20
After four tools and 87 results there are two different dates, and nothing says which one holds.
Then someone has to act on it.
Every AI tool needs its own connection, token and permissions for every system. The tangle grows, and each link fails in its own way.
- Token name
- cursor-agent
- Expires
- in 90 days
- Permission
- Write issues
- Token name
- cursor-agent
- Expires
- in 90 days
- Permission
- Write pages
- Token name
- cursor-agent
- Expires
- in 90 days
- Permission
- Repository write
Three tools, three tokens. Then the same again for Copilot, Claude Code and Cursor.
One connection per source.
Proxyma connects to each system once, and everything goes through it: you in the chat, your AI tools over MCP and your own scripts over the API.
It reads your sources, then does the work.
Proxyma indexes what you already have, then gives an AI agent the tools to work with it: reading, cross-referencing, and writing back to the systems your team lives in.
Following the safety-review skill: three open issues in PROJ reference the trap handler. PROJ-318 is the regression - it reproduces on the MX-377 bootstrap path in Cfg_Trap.c. I have commented on it with a link to the checklist and recorded it in the Safety Review Log. PROJ-402 and PROJ-455 are documentation follow-ups, so the skill says to leave them.
Type @ to mention a source or / to use a skill
AI can make mistakes. Always verify important information.
The agent searches, reads, then does the work - including writing back to Jira and Confluence.
Illustrations of the Proxyma interface with sample data. The application you install may differ in detail and continues to change between releases.
Your knowledge, and an agent that can use it.
One hub, always in sync
Office documents (.docx, .xlsx, .pptx), PDFs (scanned ones through OCR), Outlook .msg files, source code, websites, Confluence, Jira, Bitbucket and more are indexed once and kept in sync on a schedule for each type of source, so an answer cites the current version.
Sources
Files2
Web pages1
Jira projects2
Confluence spaces1
Keyword, semantic and graph search
Keyword search (BM25) catches the exact terms and semantic search (vector embeddings) finds the meaning, merged on every query and reranked if you turn it on. An optional knowledge graph links the people, teams and systems your documents name, as a map you can open.
Search
Embedding
Query expansion
Reranking
Knowledge graph
An agent that does the work
It reads the linked pages and cross-references the code, then comments, transitions the issue or opens the pull request, through your own credentials and with your permissions; by default it asks before every change, and in Read-only mode it makes none.
Private by design
It runs on your laptop or your own server: with a local model none of your documents leave the machine, with a hosted provider only the text it needs is sent to that provider, and nothing of yours reaches us unless you opt in.
- Your index stays on your own infrastructure
- Local or cloud AI, always your choice
- No account needed on the desktop app
- No telemetry without explicit consent
AI Providers
Add any service that speaks a supported protocol. Two accounts on one service are two providers.
Other integrations
What it will not pass along
An assistant with access to your systems reads a great deal that nobody wrote for it. Proxyma checks what people send, what the agent retrieves and what it answers in chat, and the checks that matter most run on your own machine, with no model and no connection.
Three checkpoints, one deployment-wide policy. The pattern checks need no AI model and no internet connection - they are on the moment you install Proxyma.
- On your machine
- Credentials, personal data and planted instructions, matched by pattern. No model, no network, nothing sent anywhere.
- Watch before you enforce
- Monitor mode runs every check and applies none of them, so you find out what a rule does to your own traffic before it does it to somebody.
- It records no content by default
- Notices, logs and metrics carry a rule name, a category and a count. Showing what a rule actually matched is a switch, off until you turn it on, and it stays on your own machine.
Pattern checks catch the attempt that was copied from somewhere, not every rewording of it. Underneath them, Proxyma never treats what it retrieves as an instruction. What the guardrails do, in full.
Illustrations of the Proxyma interface with sample data. The application you install may differ in detail and continues to change between releases.
Why not just paste the documents into ChatGPT?
- Privacy
- Your index lives on your own machine. Run a local model and no document content leaves it at all.
- Scale
- Thousands of documents searched at once, instead of the handful you can paste into a context window.
- Freshness
- A Confluence page changes or a Jira ticket arrives, and Proxyma has it. A pasted copy was stale the moment you pasted it.
- All in one index
- Files, websites, Confluence, Jira, Bitbucket, SVN, network shares and Microsoft 365
- Hybrid search
- Semantic and keyword on every query, plus a knowledge graph
- Your AI model
- Local Ollama, or bring your own cloud key
- 0 uploads
- Run a local model and your documents never leave the machine
Free for one folder. A one-time license for more.
Community indexes one folder and one website at no cost. Personal lifts those limits. Professional adds the corporate connectors your team already uses.
Try Professional free for 30 days
Community
Free forever. Everything works, at a small scale.
forever
- Semantic and keyword search, fused into one ranking
- An optional knowledge graph across your documents
- Optional query expansion and reranking, so the best passage comes first
- An AI agent that answers with citations, delegates to sub-agents, remembers your context and writes Word, Excel and PowerPoint files
- Connect your own AI model - Ollama, OpenAI, Anthropic, or any OpenAI-compatible service (vLLM, OpenRouter, your company gateway)
- Guardrails, skills, an MCP endpoint and a search API
- One local folder (100 files), one website (3 levels, 100 pages)
- Works offline with a local model - no account, no license key
Personal
Every document you own, searchable.
one-time per installation
- Everything in Community
- Unlimited local folders, unlimited files
- Recursive subfolder indexing
- Unlimited websites, crawled deeper
Professional
RecommendedYour team's Jira, Confluence and file servers too.
one-time per installation
- Everything in Personal
- Confluence (Server and Data Center)
- Jira (Server and Data Center)
- Bitbucket (Server and Data Center)
- The agent comments, updates issues and pages, and opens pull requests - by default it asks first
- SVN / Subversion
- Windows network shares (SMB)
Enterprise
Run it for your whole organization
- Cap what each person can spend on AI
- Your documents and every question asked of them stay on your own infrastructure with a local model
- Sign-in with Microsoft, Google, GitHub or any OpenID Connect provider - no second password to manage
- Roles and per-user isolation, so people see only what they should
- Index once, then share sources and conversations with your team
- SharePoint, Microsoft Teams and Outlook (Microsoft 365)
- Key accounts: review a strategic client with your team; the assistant proposes ratings and next moves, and a person confirms each one
- Everything in Professional
14 days, full refund, no reason needed. contact@proxyma.ai
A license does not include AI credits. Proxyma works with a local model you run yourself, such as Ollama, at no cost, or connects to Anthropic, OpenAI, DeepSeek, Mistral, Azure OpenAI or Google using your own account. Usage on those services is billed by them, directly to you.
Payments are processed by Lemon Squeezy, our authorized reseller and Merchant of Record. They handle VAT and sales tax at checkout, issue your invoice, and appear as the merchant on your card statement. Refunds are requested from us at contact@proxyma.ai and processed by them; see the Terms of Service.
14-day refund, anywhere in the world. No reason needed. Better still, use the 30-day trial of Professional first - it starts by itself when you install Proxyma.
Each Personal or Professional license covers one installation and is a one-time purchase - no subscription, no renewal. By downloading and using Proxyma you accept the End User License Agreement.
Common questions
Who makes this
Le Sy Hau
I build Proxyma on my own. My day work is Classic AUTOSAR software for cars, especially flash bootloaders: secure boot, secure flashing, bootloader updates and dual-bank flashing. Software that ships to hardware nobody can patch remotely teaches a particular kind of care, and Proxyma is built with it.
I wanted my own documents, wikis and repositories searchable by an AI that answers with citations, without handing any of it to a cloud run by somebody else. That is the whole product, and it is why it runs on your machine rather than mine.
Start with one folder
Free forever at small scale. No account, no card, nothing uploaded.
- No account needed - installs in minutes
- Search and chat work offline with a local model
- Your index stays on your machine
Self-hosting it on a machine with no screen?
The same app as the desktop editions, for one person with no accounts and nothing to sign in to, packaged for a home server or a VPS and reachable from the rest of your network. Paste this on the machine itself.
curl -LO https://github.com/proxyma-ai/proxyma-releases/releases/download/v0.33.0/proxyma-homeserver-0.33.0-compose.tar.gz tar -xzf proxyma-homeserver-0.33.0-compose.tar.gz && cd proxyma-homeserver-0.33.0 ./install.sh # writes .env, then stops so you can edit it ./install.sh # again, to pull the image and start
One decision to make in .env, and the installer stops so you can make it: there is no sign-in, so you say which machines may reach it. See the home server manual, or take the bundle from the release page.
Running it for a team?
Accounts, roles, per-user isolation and single sign-on are Proxyma Enterprise Server, on PostgreSQL. It is licensed and installed with us directly rather than downloaded. Write to contact@proxyma.ai, or read the Enterprise Server manual first.
Looking for something else?
Proxyma chat widget: the agent as a small chat on a website you run - tell us if you want it.
Connectors and deployments: need one we do not support, or a shape we have not thought of? Leave your email and we will get in touch.
Current version 0.33.0