Desktop is the local AI and data hub. Mobile is the pocket interface. Xiaozhi Yuntai connects AI to the physical world. NomiFun Net Infra (NomiRelay) adds self-hosted, auditable relay infrastructure for access across networks.
Desktop owns local capabilities and data; Mobile and Xiaozhi provide different interaction surfaces. When access must cross networks, you can deploy Net Infra yourself. Every project has its own source, setup path, and product guide.
01 · DESKTOP Local AI workstation
Manage models, companions, Skills, knowledge, tasks, and workflows on Windows, macOS, and Linux — then expose those local capabilities to your phone and robot.
Pocket control and collaboration
Connect to your Desktop and carry sessions, scheduled tasks, requirements, and companion status on your phone. Your own computer still does the work.
Open robot platform
An open-source pan-tilt project for Xiaozhi robots, giving a voice assistant a movable point of view and a path into NomiFun local AI capabilities.
Self-hosted network relay
NomiRelay exposes Desktop and other services behind NAT to phones and IoT devices with QUIC-first transport, TCP fallback, a web console, and explicit access policy.
If you want to quickly judge whether NomiFun fits your workflow, this walkthrough is more direct than scanning another feature list.
Open on YouTubeThe video is loaded from YouTube; open the external link if playback is blocked on this network.
NomiFun is an Agent Desktop foundation you can actually take, change, and maintain. React 19 + Tauri 2 + Rust 2024, with the desktop and web hosts sharing one frontend and backend capability set — a practical starting point for companies building their own advanced Agent Desktop.
Apache-2.0 lets individuals and companies learn from it, use it internally, fork it, and commercialize it without seeking extra permission.
Apache-2.0One React frontend and one Rust backend, shared by the desktop and web hosts. No cloud account dependency, and no core capability hidden behind a private service.
React 19 · Tauri 2 · RustWindows, macOS, and Linux are supported, so teams can adapt the system, models, permissions, and workflows to their own environment.
Win · macOS · LinuxMulti-model, multi-agent, MCP / REST, knowledge, terminals, browser use, and computer use are modular enough to trim, integrate on-prem, and keep evolving.
fork-friendlyThe desktop app already runs the full backend. Turn on WebUI and a phone or tablet can connect directly over a trusted network — with no NomiFun cloud relay, data collection, or telemetry by default.
No capability is worth peace of mind if the product quietly ships your data elsewhere. So we lay the foundation first.
Why should your chats, files and memory pass through someone else’s servers first?
Data is all-in-local, with no collection or telemetry. Network calls happen only when you explicitly configure and use a model, channel, webhook, external knowledge source, remote capability, or generation service; core workflows do not depend on a NomiFun cloud.
We cut several advanced, fun features — the open-source build ships without them for now — because the data safety of every user and every developer comes first. Those capabilities return once they are mature and controllable.
See the Roadmap for what’s nextYour device
SQLite as the single source of truth · isolated work dir per session
Everything lives on your machine: SQLite is the single source of truth, and every session gets its own isolated working directory.
all-in-localFully open-source under Apache-2.0. The code is public and open to audit — how data flows, and whether it leaves, is all visible.
Apache-2.0Free for commercial use. No ads, no membership, no paywall — safe for individuals and enterprises to rely on long-term.
no ads · no membershipThe more you use it, the better it knows you.
Three built-in characters — Mochi the mochi bunny, Ink the amber-eyed cat, and Bolt the hovering robot, all hand-drawn in code; or design any companion IP you like — even your own family or pets. Plus persona presets (Lively / Calm / Sassy, or free text); behavior capture is mostly off by default and opt-in, distilled by an LLM into long-term memory — so companions learn you over time.
Three built-ins — Mochi the mochi bunny, Ink the amber-eyed cat, Bolt the hovering robot; or customize any companion IP you want — even upload your own family or pets to make a one-of-a-kind companion.
Memory supports shared and per-companion private scopes: chat memories default to the active companion, learner output is shared. Skill libraries remain isolated per companion.
It auto-mines reusable skills from tool-call sequences (parameter values excluded) → generates a reviewable SKILL.md, enabled only after it reviews it with you.
Gift a skill from one companion to another and turn on shared learning. Teach just one companion, and let it teach the rest.
Soft Japanese · mochi bunny
All done! What else should we tackle — just say the word~
Persona preset
Pick a companion
auto-mined from tool-call sequences
A reviewable SKILL.md is generated — enabled only after you confirm
Anywhere, anytime — with just a network and a chat app, command your far-away companion to operate your computer.
Every companion is a full, independent individual that can connect to 12 built-in IM channels on its own. Send orders from any chat app and it gets the work done on your computer — one bot bound to one companion.
Telegram
Lark / 飞书
钉钉
微信
企业微信
Slack
Discord
Matrix
Mattermost
Twitch
Nostr
QQ BotTelegram, Lark, DingTalk, WeChat, WeCom, Slack, Discord, Matrix, Mattermost, Twitch, Nostr, and QQ Bot are compiled into the default Desktop build.
Issue commands from any chat app; the companion does the work on your computer — instant control from anywhere.
Each companion is a complete, independent individual that fully drives the whole system; one bot binds to exactly one companion.
Switch on the WebUI on your LAN / office network / VPN, and your phone or tablet gets the exact same interaction as the desktop. Pull out your phone and keep working, anywhere, anytime.
The experience on phone or tablet is identical to the desktop — not a cut-down lite version. Same interface, same full set of capabilities.
On a LAN / office network / VPN, scan once to connect: a one-time token valid for 5 minutes, a Host/Origin allow-list that blocks DNS-rebinding, and never through any social platform.
Want to keep using Codex / Claude? The platform’s terminal sessions (PTY, presets Shell / Claude Code / Codex / Gemini) give you a native, uninterrupted remote session.
Scan to connect this device
One-time token · 5-min TTL
One NomiFun — authenticate on another device and continue in the full workspace 0.0.0.0:25808 · LAN direct
$ claude
running in NomiFun terminal — session kept alive
context restored · attached to the same working directory
editing src/lib/router.ts … wrote 3 edits
$
Drop a requirement on the board and it takes over: smart per-tag rotation, single-loop execution, with lease sweeping and boot resume keeping it reliably alive — you just direct, it works for you.
Pending→InProgress→Done/Failed/NeedsReview, rotated per tag with single-loop execution. The backend is the single source of truth.
A lease sweeper re-pends orphaned tasks every 60s, and boot resume picks work back up after a restart — dependable even when no one is watching.
Rule tier auto-retries / backs off / nudges / auto-confirms read-only permissions; the sidecar model tier handles genuine decision stalls and can trigger a model failover queue.
Snapshot competitor pricing
Roll up weekly metrics
Sweep session health
Fix login regression
Draft release notes
NotifiedYou just direct — it works for you, reliably.
Has everything, works with everything, plays well with everything.
Platform capabilities are exposed through three public facades: /mcp (full Remote surface), /mcp-agent (a curated do-work subset), and /v1 (REST + OpenAPI 3.1 + SSE streams). Let Claude, Codex, or any agent drive NomiFun directly over MCP/Skill.
/mcp exposes the full Remote surface, /mcp-agent a curated do-work subset, and /v1 offers REST + /v1/openapi.json (OpenAPI 3.1) with SSE streaming responses.
Bearer tokens (stored as SHA-256, compared in constant time) authenticate calls; the caller acts as that companion on the Remote surface. Headline tools: nomi_agent_run (streaming) + nomi_agent_result (polling).
DangerTier × Surface × Decision finely govern every call; knowledge writes are additionally controlled by Disabled / Manual / Auto policy.
POST https://<host>/mcpAuthorization: Bearer <companion-token>Content-Type: application/json { "method": "tools/call", "params": { "name": "nomi_agent_run" }} # Streamable-HTTP · 全量 Remote 能力面MCP Streamable-HTTP · full Remote capability surface
Knowledge write-back is not a universal IM staging rule: Disabled is read-only, Manual writes only when you explicitly ask, and Auto may extract durable knowledge at turn end. External IM channels also require channel_write_enabled; when enabled, writes land in the knowledge-base body.
Centrally manage, use and track the knowledge scattered across your system — and configure knowledge bases, skills, agents, MCP and models once to reuse them across any session, agent, companion or terminal.
Knowledge scattered across files, web pages and chats — hunting it down every time? Pull it into one manageable, trackable knowledge hub.
Manage, use and track your knowledge bases from one place: curated markdown directories, live URL snapshots, and policy-controlled write-back, with Feishu as a source — mounted on demand into any session, agent, companion or terminal. The same “config once, use anywhere” applies to skills / agents / mcp / models too.
Disabled turns write-back off. Manual skips turn-end extraction and writes only when you explicitly ask the Agent to record something. Auto may retain durable knowledge at turn end. Qualifying writes land in the knowledge-base body, with append and compare-and-swap updates for existing documents.
A knowledge base is a markdown directory you curate, mounted into the session workspace so it’s instantly queryable. Bind it to a companion or a session.
Turn web pages into knowledge entries: Live fetch or pinned Snapshot, with built-in SSRF protection and a headless browser for JS-heavy pages.
Choose Disabled, Manual, or Auto per mount: off, explicit-request only, or Agent-directed retention. Existing documents use append and compare-and-swap semantics to reduce overwrites and collisions.
For browsing and computer control we don’t bolt on someone else’s runtime. It’s self-developed, in-process, and served to the model as native tools.
Spawning a Playwright / Node subprocess is slow, heavy and hard to govern — can it be cleaner?
A self-developed in-process Rust engine and native computer-use, served to the model as native tools: faster, fewer tokens, fine-grained control, and source you can extend.
Bolted-on runtime (the usual way)
NomiFun native
An in-process Rust CDP browser engine, Chromium-only, with no Playwright and no Node subprocess — about 32 actions drive the page directly.
Screen capture + input synthesis, backed by each platform’s accessibility engine: macOS AX / Windows UIA / Linux AT-SPI — about 21 actions that truly operate your desktop.
A three-axis capability gateway (danger × surface × decision) gates every tool call — allow / confirm / deny by policy; open source, extensible by developers.
Installed means ready — no agonizing over "which agent do I install first".
The built-in nomi agent needs no extra install (the `nomi` CLI binary) — usable the moment NomiFun lands. Four native backends cover the major clouds, OpenAI-compatible reaches 26+ providers / presets, and ~19 external ecosystem agents work alongside it — all fed models and local capabilities by NomiFun.
$ nomi The `nomi` CLI binary ships with NomiFun and works out of the box — no separate download, no extra environment to wire up. Once installed you can start conversations and delegate tasks right away.
Anthropic 
OpenAI 兼容 
Amazon Bedrock 
DeepSeek 
Gemini 
Qwen 
Kimi 
Zhipu / GLM 
StepFun 
Doubao / Volcengine 
Mistral 
xAI 
OpenRouter 
New API 
ModelScope 
InfiniAI 
Novita 
PPIO The OpenAI-compatible backend + a custom base_url reaches DeepSeek / Gemini / Qwen / Kimi / GLM / StepFun / Volcengine / OpenRouter / New API / Ollama / vLLM / Azure and more.
Not just the built-in nomi agent — Claude Code, Codex, OpenCode, Nanobot, OpenClaw, and more can connect, with NomiFun supplying their models and local capabilities.
Claude Code Codex CLI Gemini CLI Qwen 
Droid 
Goose 
Auggie 
Hermes Kimi 
OpenCode 
Nanobot 
OpenClaw Beyond the in-house nomi agent, every agent above is supplied models by NomiFun — one provider config, used everywhere.
Open an interactive session and chat back and forth with the agent, or spin up a real PTY terminal directly. The terminal ships with handy presets so you land in a familiar workflow.
$ nomi
› ready · interactive session + pty
$ nomi run --stream
NomiFun goes beyond text tasks. Interactive Agent sessions and real PTYs handle execution, Creative Workshop arranges image, video, audio, and generation nodes on an infinite canvas, and mini apps turn a session artifact into a local tool you can keep iterating.
Multi-model, multi-agent collaboration, execution graphs, previews, and working directories stay connected from discussion through delivery.
A real in-app PTY for Shell, Claude Code, Codex, Gemini, and other workflows, with the same knowledge and local capabilities available inside.
Arrange images, text, video, generators, and speech nodes on an infinite canvas for nonlinear visual, audio, and video creation.
Create a single-file mini app from a normal session, then materialize a working copy from the library to iterate, import, and publish instead of losing it as a one-off answer.
The old site captures have been replaced with new screenshots from the installed Desktop and a Mobile client actually connected to it.
This is not a future-feature wish list. The 2025 capabilities were built in-house and put into real internal use; the early-2026 innovations have shipped. Open-source scope, public entry points, and internal deployment can differ, while unreleased confidential work is explicitly identified.
2025
Built and in internal production use
These six capabilities were implemented in 2025 and used in real projects and daily internal work. They are delivered systems, not forward-looking concepts awaiting implementation.
Early 2026
Product innovations shipped
By early 2026, NomiFun had expanded beyond a standalone workstation into mini apps, direct mobile access, Agent clusters, customer service, and an IM Channel gateway. These entries describe shipped capabilities, not a gradual future rollout.
NomiFun began as an experimental tool for our own work and a few content-creator friends. We opened the source after the 2026 UI rebuild. Bandwidth is still limited, and the project remains both experimental and public-interest oriented, so we would rather explain the code, boundaries, and real progress than push hard marketing.
See how to contributeNomiFun is built part-time and is in its pre-1.0 growth phase. Whether you write code, run communities, or love to share — there is a place for you here.
Right now we really need these roles to grow the project
Open PRs, fix bugs, sharpen capabilities. Every merge counts toward your contribution score and keeps your name in the project.
Answer questions, organize discussions, curate docs and best practices — turning scattered enthusiasm into a sustainable community.
Write posts, give talks, record videos — introduce NomiFun to the people who need it and help good work get seen.
Individuals and enterprises alike are free to modify and commercialize NomiFun, free of charge — that is your right under Apache-2.0, no permission needed from us.
If you do ship it commercially, we would love a quick note. Not for licensing — just that, as a developer, a word of recognition from real-world use means the world. A little positive feedback goes a long way.
Any legal risk from modification or commercial use is borne by the user.
If installation, model-provider configuration, WebUI, or MCP usage is getting in your way, reach out through GitHub or email. We prefer turning reusable answers into issues, docs, and community guidance.
Note: NomiFun is still growing. Complex deployments, third-party service integration, and compliance risk need to be assessed by the user.
Contact us
