Your private interface
for every AI model.
PUI.ai runs in your browser, keeps workspace data on this device, and lets you choose the provider, model and tools for each conversation.
Quick start
A useful first conversation takes less than a minute.
Choose a model
Open Models or use the model control below the composer. The public OpenRouter connection shows free models; your own key unlocks its full compatible catalog.
Start a chat
Ask directly, pick a starter, reuse a prompt, or press / for workspace actions. Responses stream into the active conversation.
Organize your work
Drag conversations into colored folders and nested folders. Projects add reusable agents, models and local knowledge.
Conversations
Chats are durable browser-local workspaces, not server accounts.
Folders, colors and selection
Create nested folders, then drag chat rows, colored working tabs or folders into place. Select filtered chats for bulk organization, or move every active and archived conversation to recoverable Trash after a count-based confirmation.
Summary and title
The summary control creates a compact topic for the active chat. Keep it beside the title, use its suggested title, regenerate it, or remove it.
Saved or temporary
Open the composer + menu for a normal New chat or a browser-tab-only Temporary chat. Temporary content stays out of conversation history, local search, usage reports and backups unless you explicitly choose Save chat; reloading or closing that browser tab discards it.
Tab-local follow-up queue
While an answer is running, queue more messages for this chat in this browser tab. They run automatically in order only while the tab stays open; they are not saved or synced. PUI asks the browser to warn before a reload or close while items are waiting, and Stop remains available.
Search and saved views
Search titles and message text locally. Filters, folder scope, tags and sort order can be saved as reusable conversation views.
Models & providers
You decide where each request runs.
Workspace data stays local, but a message leaves the device when you send it to a cloud provider. Review the selected provider before sending sensitive content. Local Ollama, LM Studio and oMLX endpoints can keep inference on your own computer.
Single model
Use one selected provider and model for a normal multi-turn conversation.
Agents
Save an instruction, starters, prompts, skills and carefully permissioned tools as a reusable assistant.
Compare
Ask several models independently, review their answers side by side and choose the primary lane.
Council
Run anonymized candidates and let a selected judge synthesize a final answer with visible evidence.
Files & knowledge
Add context without turning pui.ai into a file-hosting service.
- Composer files: drag images, PDF, DOCX, TXT, Markdown, CSV and other supported text formats into the chat. They are processed locally and attached only to the next request.
- Vision: locally sanitized PNG, JPEG and WebP images are sent only when every selected model supports image input.
- Asset Vault: save normalized uploads and sanitized images for reuse. Supported files returned by an approved MCP call are verified and saved here automatically with their response.
- Answer downloads: save fenced Markdown, JSON, CSV, HTML, SVG, Mermaid and source code locally. MCP images and documents appear as response cards with preview, a two-step Save or open action and Asset Vault access.
- Knowledge: index bounded local passages for a project, inspect retrieval results, and decide which context an agent may use.
MCP & tools
External actions stay visible and permissioned.
Exa Search
The reviewed Exa Streamable HTTP server starts enabled in new workspaces for live public search and page retrieval. You can remove it completely when another MCP server already provides search, and deliberately add the reviewed template again later.
Verified per browser
A saved catalog stays inactive until this browser tests the current endpoint, credential binding and tool schemas. A protected transfer can preserve exact approval preferences, but keeps every allow paused until that fresh destination-device test succeeds.
Readiness in chat
The compact composer status shows one truthful next step: enable, unlock, test, grant context, assign, choose a compatible direct model, or use the verified tool count. It never performs a tool call or widens permission by itself.
Approval choices
Approve once, always allow the exact tool, deny once, or always deny it. Persistent decisions are scoped to the exact server and command.
Approve once for every chat
One global MCP action can permanently allow the full reviewed catalog of every ready server in this browser workspace. It lists included and skipped servers, keeps exact denials stronger, and never silently includes new or changed tools.
Files from tools
Supported MCP images, text, PDF and DOCX outputs are verified outside the model transcript, saved in this browser and shown directly in the answer. HTTPS resource links are never prefetched.
Safe portability
Workspace imports never inherit hidden tool authority. Imported persistent decisions and custom-server grants are cleared for review.
Privacy & storage
The browser is the primary data boundary.
Stored in this browser
- Chats, folders, projects and preferences
- Prompts, agents and model metadata
- Local Knowledge indexes and Asset Vault content
- Optional encrypted browser credentials
Not stored by the pui.ai server
- Your chat database or local files
- Your provider and MCP credentials
- A shared server-side conversation account
- Analytics profiles or advertising identifiers
Use Settings → Move to another device for recovery or a browser/device move. A preflight inventory proves which provider/local-runner/MCP credentials enter the password-protected .pui-backup; restore validates and re-encrypts each one in the destination browser. Exact MCP approval preferences may transfer as paused preferences, becoming usable only after one matching destination-browser test. Local Network permission, private-CA trust and local runtimes remain device-specific. Plaintext and ordinary Workspace JSON remain credential-free and never restore imported approval authority.
For one conversation, use Export chat → Secure hand-off → Protected package. The browser creates an encrypted .pui-chat without credentials, instructions or files. Send the file and passphrase through different channels. The recipient uses Settings → Receive secure chat hand-off, reviews the local preview and explicitly chooses Import clean copy; nothing merges automatically.
URLs & deep links
Public documents and private workspace views have stable browser routes.
/Generic workspace entry; opens the active browser-local conversation/chat/<chat-id>Opens that conversation in the active browser workspace/workspace/modelsModel catalog and provider connections/workspace/mcpPrivate MCP setup, device tests, assignments and approvals/workspace/<section>Projects, Knowledge, Agents, Prompts, Usage and Settings; browser Back and Forward restore the view/featuresProduct overview for conversations, models, files and workflows/local-aiOllama, LM Studio and oMLX setup and connection architecture/mcpRemote MCP tools, built-in Exa search and approval boundaries/businessBusiness support, maintained self-hosting and Enterprise/pricingFree use, voluntary support and commercial deployment options/releasesCurrent version, changelog, compatibility and verification notes/docsHelp and product documentation/privacyDetailed data-flow and privacy disclosure/securitySecurity architecture and trust boundariesPrivate workspace URLs carry no workspace data. A conversation URL is not a public share link, and a /workspace/… URL contains no workspace ID, key, token, filter or message. Routes resolve only against the currently active browser-local workspace; another browser or device sees its own data.
Keyboard
Shortcuts remain available without occupying composer space.