TACHI_STUDIO — every AI model. one window. zero terminal.

01

Free by default

Start with 17 free upstreams behind the built-in FreeLLM router — no card required. Bring your own Anthropic or OpenRouter account when you want it, or run fully local. One router drives them all.

Pollinations FREE LLM7 FREE Groq FREE Cerebras FREE SambaNova FREE NVIDIA NIM FREE Cloudflare FREE Gemini FREE Mistral FREE OVHcloud FREE Scaleway FREE Cohere FREE GitHub FREE Zhipu FREE Ollama Cloud FREE Kilo FREE OpenRouter FREE

FreeLLM ROUTER Anthropic OAUTH OpenRouter OAUTH OpenGateway FREE free-claude-code FREE Bankr PAID Venice PAID Surplus PAID imgnAI PAID Ollama LOCAL llama.cpp LOCAL

~1B FREE TOKENS / MO 0 API KEYS 0 CARDS

02

Smart routing

A local, sub-millisecond classifier — no LLM in the loop — sorts every prompt into difficulty × task type; a bandit re-ranks models by what actually works, with cooldowns and failover. Leave it on AUTO. Fire a prompt and watch it pick.

CHAT — SMART ROUTING

SMART Bankr Venice Surplus SAVED 0%

TACHI_STUDIO — CHAT REAL CAPTURE

The real Chat tab — searchable history rail, provider picker, streaming composer.

Compare

Race the models

2–6 models answer the same prompt in parallel columns — live tok/s and time-to-first-token per column. Local models join keylessly against the cloud.

Fusion

Panel + judge

A panel answers, an arbiter delivers one reply — compare, synthesize, best-of-N or majority. FRONTIER and BUDGET presets, one chip in the composer.

Budget guard

Spend, capped

Every call lands in a 30-day ledger under a hard monthly cap. Leave the router on AUTO, COMPACT a long chat without losing it, and watch the live tally of TOKENS SAVED and PROMPT-CACHE hits routing just earned you.

03

Go private

One switch (Cmd/Ctrl + K). Cloud providers lock, agent harnesses refuse to start, network tools are denied at the main-process boundary — and the same router points at models on your own machine. Pull the cable, it still answers.

CLOUD PRIVATE

[ PRIVATE ] — ZERO EGRESS

ProviderFreeLLM[ LOCKED ]

ProviderBankr[ LOCKED ]

ProviderVenice[ LOCKED ]

ProviderSurplus[ LOCKED ]

LocalOllama

Localllama.cpp

YOUR MACHINE ✕ CUT INTERNET EGRESS: OPEN

WEBFETCH ✕WEBSEARCH ✕CURL · WGET · SSH ✕MCP FETCH → LOOPBACK ONLYOLLAMA ✓LLAMA.CPP ✓

TACHI_STUDIO — PRIVATE REAL CAPTURE

One switch, and every cloud path leaves the app — it tells you exactly what it turned off.

04

One click to any model

The Catalog reads your hardware, tells you what fits, and installs with one click. Every artifact is SHA256-verified against the publisher’s digest — packaged builds fail closed on anything unpinned. No terminal, no GGUF archaeology.

RAM 94GB FREE GPU RTX 3080 Ti VRAM 12GB CPU 24 CORES

SEARCH ›

gemma-4-12b

12B · Q4_K_M · 7.2 GB

FITS IN GPU

CHATVISIONCODE

IDLE — 7.2 GB

qwen3-coder

32B · Q4_K_M · 19.8 GB

PARTIAL — CPU OFFLOAD

CODETOOLSREASONING

IDLE — 19.8 GB · also: VISION sibling

deepseek-r1

70B · Q4_K_M · 42.5 GB

TOO LARGE

REASONINGCHAT

NEEDS 48GB VRAM — route it to cloud instead

05

Wire your own AI

Providers, agents, roles — plus text, reference-image and media tiles — patch together like hardware. Every run drops a result card you branch from; flows save to disk and reopen from the Flows rail.

PROVIDER SRC

OpenGateway

nemotron-3-ultra :free

streamtools

AGENT MAIN

OpenClaude

model — upstream provider

RUNPIN

PERMISSION ACL

Generalist

full toolset — default

readwritebashwebfetch

▸ RUN ALL (IN ORDER) — RESULT CARDS DROP LIVE

TACHI_STUDIO — NODES REAL CAPTURE

The real node canvas — a media pipeline wired from reference image to video.

Pin + cache

Re-run only what changed

Pin a node and RUN ALL reuses its cached output instead of paying to run it again. Every run still drops a result card you branch a fresh node off.

Cost + errors

See spend, catch the fail

Every node wears a live cost chip and the canvas totals against your budget cap. Wire a node's red error handle and a failed run branches down it while the happy path just skips.

Save + run

Save it, run it all

Graphs persist as .tachi-flow.json — reopen from the Flows rail, auto-layout, and RUN ALL executes the whole canvas in order.

STICKY NOTES CARRY THEIR NODES SUBFLOWS COLLAPSE TO ONE NODE THE GRAPH TRAVELS INSIDE THE PNG FLOW DOCTOR — ONE-CLICK REPAIR DRAG IT FROM THE PALETTE

06

Built for artists

Chain nodes into a media pipeline: one reference image, a vision model that writes the scenario, and image-to-video that keeps your character on-model — every scene, every style.

REFERENCE IMAGE INPUT

your character — one image is enough

→

PROMPT MAIN

› make scenario for 5 sec promo with my character
› vision — reads the reference, writes the shots, keeps the crab on-model

Bankr · Claude Opus 5 — vision

→

MEDIA — VIDEO OUTPUT

omni — image-to-video · real render · 5s

Bonk the FUD — generated scenario — Bonk the FUD ON-MODEL

GM wave — generated scenario — GM Wave ON-MODEL

Saint of Shipping — generated scenario — Saint of Shipping ON-MODEL

Base Mode — generated scenario — Base Mode ON-MODEL

PROMPT → IMAGE → VIDEO ONE CHARACTER, EVERY SCENARIO NO TIMELINE APP NEEDED

07

Describe it. Ship the film.

The Design tab is a motion-graphics studio: describe a launch film, a poster, a what's-new reel — the model writes a real composition and it plays live in the app. Then export it, voice it, or hand it your brand.

Prompt → film

Three engines

HTML, Remotion and HyperFrames compositions in a live player you scrub with the cursor to land any frame — generation self-heals on truncation and repairs its own errors.

Export

MP4 · PNG · PDF

H.264 video at 1080p, 2K or 4K, stills and print-ready PDF — rendered by a managed Chromium the app installs itself. No timeline software.

Voice + sound

Voiced, locally

Narration via local TTS — studio-quality Kokoro or fast Piper — plus sound in animate mode. Nothing uploaded to voice a film.

Brand-from-URL

Point at a site

Give it any URL — palette, typography and tone are extracted and applied — or attach your own footage, stills and folders as reference. On-brand from the first frame.

Edit both ways

Canvas or code

WYSIWYG editing right on the render, or the full source in a Monaco editor — same composition, your choice of handle. The history rail keeps every take, and brand plus attachments are remembered per project across sessions and restarts.

Audio overview

Doc → podcast

Turn a repo, a document or a chat into a two-voice audio overview (~90 s) — scripted and synthesized entirely on your machine.

08

Set it going. Walk away.

The app improves its own source, installs its own tools, and fires jobs while you sleep — then tells you the truth about how it went.

TACHIAPP

It edits its own source

A pinned row in the sidebar opens the agent already bound to the app's own code — zero folder picking. Ask for a change, read the diffs, ship it. The Comic theme in this build was written that way.

MCP marketplace

24 servers, one click

A curated catalog in Settings → Connections. Fill the two fields a server actually needs, hit INSTALL — no JSON to hand-edit. Secrets land in the OS keychain, and each server is egress-classified so local-only ones keep running in PRIVATE mode.

Scheduler

The 2am job fires

Settings → Advanced → SCHEDULED runs a saved flow or a prompt once, daily, weekly or every N minutes — offline, no cloud cron. Sleep through an occurrence and a per-job missed-run policy decides: catch up on wake, or skip. The spend cap is checked before every fire.

Honest endings

It admits it stopped

When a run ends without finishing, the harness spends exactly one automatic nudge to continue — and if it still stops short, an amber ENDED WITHOUT COMPLETING badge says so, with a CONTINUE button. The run log records it as incomplete. No checkmark on a give-up.

Resilience

RETRYING 3/10

A dropped stream is a blip, not a dead run: up to ten reconnects with exponential backoff and jitter, the server's Retry-After honored, and a live banner in both chat and the agent. The round only replays while no tool has run, so nothing executes twice.

Operator powers

Fan out. Or loop.

spawn_agents hands work to bounded child agents in parallel and folds their answers back, each child's trail tagged in the transcript. /loop repeats a prompt or a command on an interval behind a LOOP chip — and paces itself when you don't name one.

SOURCES CHIPS UNDER GROUNDED ANSWERS MEMORY FACTS YOU CAN EDIT YOUR OWN LAN ENDPOINT FAST / BALANCED / CREATIVE MOVE MODEL WEIGHTS OFF C: FIVE THEMES — ONE WRITTEN BY THE AGENT

09

Inside the workstation

One window, every tab a tool. Slide through the racks.

CHAT01 / 09

YOU
summarize this repo and draft a README

ROUTER · TOP · claude-opus-5
Scanning 142 files… here's the structure: a desktop studio in Electron with a provider router at its core…

YOU
now in Russian

Chat — streams from any provider; folders with per-folder system prompts, full-text history search, a slash-command layer in the composer, and attach-a-folder RAG: a local index grounds the answer in your files and prints the SOURCES chips it used.

AGENT02 / 09

AGENT · OpenClaude
$ npm test
✕ 3 failing → reading stack traces…

PERMISSION · Generalist
read ✓ write ✓ bash ✓ webfetch ✓

HARNESS
TACHI · Codex · DarkSol · OpenClaude — PLAN / BUILD · THINK / ULTRA
TRUST · SAFE / STANDARD / AUTO

Agents — TACHI, the first-party harness, plus Codex, DarkSol and OpenClaude — each fenced by explicit permissions and a SAFE / STANDARD / AUTO trust preset.

MEDIA03 / 09

IMAGENano Banana 2

VIDEOqueued · 2 jobs

MUSICidle

TTS / STTPiper · Whisper

Media — image, video, music, speech via Venice, Surplus or imgnAI — or fully local with stable-diffusion.cpp, Piper, Kokoro and Whisper.

DESIGN04 / 09

YOU
45-second launch film — brand it from my site, voice it

DESIGN · REMOTION
rendering scene 3 / 6 — H.264 · 1080p60
voiceover: kokoro (local) ✓

EXPORT
MP4 ✓ PNG ✓ PDF ✓ — up to 4K · history rail keeps the session

Design — a motion-graphics studio: prompt → animated, voiced film; edit on-canvas or in Monaco; export MP4/PNG/PDF up to 4K.

CATALOG05 / 09

gemma-4-12bFITS IN GPU

42% · 5.1 MB/s

qwen3-coderPARTIAL

19.8 GB

Catalog — hardware-aware model store. Fit tags, live download rate, Run.

API + MCP06 / 09

LOCAL API · 127.0.0.1:11435/v1
$ curl -H "Authorization: Bearer tachi_…" /v1/chat/completions
{"choices":[…]} — OpenAI-compatible · SSE streaming

MCP SERVER
fs / git / llm tools → Claude Desktop, Cline, Codex
token-gated · read-only by default

API + MCP — a drop-in OpenAI endpoint for any SDK, an in-app MCP server external agents can borrow, and a 24-server MCP marketplace that installs with one click.

WALLET07 / 09

NETWORKBase (EVM)

KEYSOS keychain — never cross IPC

SENDexplicit confirm required

TOKEN$TACHI

Wallet — an EVM wallet on Base (ethers v6); keys live in the OS keychain, every real transaction asks first.

SWARM08 / 09

BACKEND ENGINEER
API schema drafted → handing to frontend

FRONTEND ENGINEER
building the settings panel…

SECURITY ENGINEER
audit: 1 finding — unscoped token in config

Swarm — role agents coordinated over plain git (gnap), each with its own leash — no server needed.

LEARN09 / 09

GUIDED TOURSDesign · Catalog

FIRST LAUNCHone-time hero

Learn — a fresh install opens on a one-time Learn hero that walks you through the studio; returning users skip straight back to work. Step-by-step tours light up the Design and Catalog tabs on demand.

TACHI_STUDIO — CHASSIS REAL CAPTURE · MOTION

Bankr, OPUS-5, TK-05 — one running session, three chassis. Real geometry, not a palette swap.

10

Everywhere, every language

Eight languages, one UI. Windows ships today; the macOS build is next, Linux is planned. Made for normies and pros alike.

SETTINGS — LANGUAGE

› New chat — choose a model

ENESFRDE RU中文日本語한국어

Windows.exe — per-user NSIS, no admin / UACTODAY

macOS.dmg — Apple Silicon + IntelNEXT

Linux.AppImage / .deb / .rpmSOON

11

No gatekeepers

The code is open, the router is yours, and the community holds the keys.

Open-source

Read the code

MIT-licensed, top to bottom — the router, the harnesses, the canvas. Electron 43 · React 19 · TypeScript. Fork it, audit it, point it at your own gateway.

Download for Windows → Star on GitHub →

Community

$TACHI

The community token behind the build. Follow development, vote on the roadmap, be early.

Follow on X →