A completely uncensored, locally-run AI coding assistant and agent workspace powered by Ollama (with optional LM Studio and cloud-provider models). No refusals, no guardrails, no limitations — the model will assist with any request without hesitation. Includes a full mobile app development studio: analyze an app concept and plan its feature set (optionally grounded in a real decompiled .apk for interoperability/prototyping), generate a complete project from that feature list, run it live in-browser, and build a real, installable Android APK — entirely offline, without leaving the app.
Built by CopperArch.
Licensed under the PolyForm Noncommercial License 1.0.0 — free to use, modify, and share for any noncommercial purpose. Commercial use (selling it, running it as a paid service, bundling it into a paid product, etc.) requires a separate commercial license — open an issue on this repository to arrange one.
Chat history now lives in SQLite instead of a flat JSON file — each conversation pins to one persistent session instead of drifting into duplicates, and a page reload correctly returns you to the conversation you were actually using instead of an arbitrary one. Batch conversation delete (right-side checkboxes, delete-selected or delete-all) removes conversations from the database, not just the page.
Added OpenCode Zen (several genuinely free models, including a strong stealth-labeled coding model) and Groq (fast open models, generous no-card-required free tier) as selectable cloud providers alongside OpenRouter, NanoGPT, and haimaker.ai. If a free model hits a shared-pool rate limit, the app now retries and, if needed, automatically falls back to a different free model on the same provider instead of surfacing a raw error. Also added /api/models/rankings — a coding-focused ranking across local, free, and paid models refreshed against live signals rather than a hand-maintained guess.
The Image Generation tab now does actual image generation instead of describing an imagined image in text: OpenRouter-backed generation/editing, or a fully local GPU pipeline with automatic hardware detection (NVIDIA/AMD/Apple Silicon) and an isolated environment so its dependencies never collide with the main app.
A light/dark theme toggle, a collapsible off-canvas sidebar for narrow screens, full-text conversation search, a Ctrl+K command switcher, and edit-and-resend / regenerate controls on any message.
Long conversations no longer silently truncate — as a chat approaches a model's real context window, older turns are automatically folded into a compact summary instead of being dropped, with no setting to remember to enable.
The agent's tool-calling path was rebuilt around a typed tool-registry interface for cleaner dispatch and easier extension going forward, with the same tool set (file edit/read/write, search, code execution, shell commands, subagent delegation) behaving identically from the user's side.
- A backend shutdown/restart race could leave a child process unaccounted for, occasionally causing a hard crash on restart; shutdown now waits for every child process to actually exit.
- The model picker could silently revert your selection back to a local model mid-session during a routine background refresh.
- A model-picker bug (affecting every cloud gateway provider) could send an invalid, undefined model reference instead of the one you actually picked; fixed at the source.
- Cloud provider errors (e.g. a rate-limited free model) were shown as a raw provider error dump instead of a clean message.
The chat bar now has a command button row (New Chat, Compact, Skills, Agent, Model, Clear, Help) plus full slash-command autocomplete — type / and pick from the list with the arrow keys, Tab, or Enter. /compact summarizes everything except the last few turns into one dense summary message, freeing context on long sessions; /agent <task> jumps to Agent mode with the task pre-filled. Conversations get automatic titles generated after their first exchange, and a toast announces it whenever a session teaches the agent a new auto-learned skill.
- Plan tracking —
todo_writegives the agent a live task list it must seed before multi-step work and update as it goes; the chat renders it as a progress checklist (◐in-progress,✓done) inside the reply. - Surgical edits with diffs — new
edit_filetool (exact old-string → new-string, uniqueness enforced, near-miss recovery hints). The agent is instructed to prefer it over whole-file rewrites; every edit/write returns a unified diff rendered red/green in the chat. - Content search — new
grep_filestool (regex over file contents withfile:lineresults, skipping junk/binary dirs), alongside the existing filename search. - Subagents — new
tasktool delegates self-contained research/exploration to a nested agent with a fresh context window; only the distilled report comes back (one nesting level, no sudo inside).
Skill distillation now fires on file edits/greps/subagent runs (not just web/code), rejects near-duplicate lessons by name/description similarity, and caps the auto library at 120 (evicting oldest) so the system-prompt skill directory stays small.
A 🤖 Agent Suggestions button analyzes the currently open project and proposes its highest-value next tasks as one-click cards — each is a fully self-contained agent instruction (real files, real functions) with a ▶ Run in Agent button that switches to Agent mode and starts it immediately.
When a cloud model is selected, every request now uses the standard token-saving stack: prompt caching (Anthropic explicit cache breakpoints on the system prompt and conversation prefix — the big one, ~90% off cached input; OpenAI/Gemini cache automatically and are read back the same way) and history slimming (older tool outputs are elided from re-sent history instead of paying for them again every turn). Every cloud call's exact token counts — including cached tokens — are priced against a built-in rate table and logged to a monthly spend ledger. The chat shows the estimated cost of each turn right after the token count, plus a running "💳 $X of $Y budget · N% left" line (budget configurable from the token-usage ⋮ menu, default $20/month, resets each month, colors warn at 35%/10% left). All costs are marked ≈ — awareness, not an invoice.
Build Android now offers a Local (Debug), Local (Release), or EAS Cloud option. Local builds run entirely on your own machine via the Android SDK + Gradle — no Expo account, no cloud service, nothing leaves your computer. Release builds are signed with an automatically-generated, persistent local keystore, so they're installable straight away. The one-time SDK/JDK setup is auto-detected; the app tells you which build methods are actually available on your machine.
A toggle that switches the App Analyzer from "plan a competing app" to "replicate an existing app with full feature parity" — exhaustively inventories the real app's features instead of proposing a cut-down MVP, then hands the feature list straight to the Project Generator with one click.
Drop an .apk (or scan your drives — home folder and mounted external drives — for one) to ground Clone Mode in a real decompiled app: package/version/permissions from the manifest, plus a filtered sample of real in-app UI strings pulled from the compiled bytecode, all fed into the analysis as ground truth alongside the model's own knowledge.
Generate Project now has a Run Code button: saves the project, npm installs it, guarantees the web-support packages a generated app routinely omits (react-dom, react-native-web, Metro/webpack runtime depending on Expo SDK age), and starts a live browser preview — embedded directly in the panel via an iframe, with a Stop button. Includes an automatic dependency-repair loop: local models occasionally invent a package version (or a whole package) that was never published, so a failed install is parsed for npm's specific error and retried with the offending entry relaxed or removed, rather than failing the whole run over one bad line the model wrote.
Build a real, installable .apk via EAS Build (Expo's free-tier cloud build service) — no local Android SDK needed. Paste a free Expo access token once (validated live against the real account), and it configures, builds, and downloads the finished APK straight into the project folder.
Reload any previously saved project back into the Generate tab — Save to Disk, Run Code, and Build APK all work on it exactly like a freshly generated one.
- The App Analyzer/Project Generator's real token usage wasn't being added to the weekly usage counter at all; the Agent tab wasn't tracking tokens either, and the Chat tab was silently double-counting them. All four are now correct.
- Chat's date/time questions could get answered with a hallucinated date from the model's training cutoff instead of the real one — pure date/time questions are now answered directly from the system clock, and the chat system prompt always carries the real current date/time as grounding for everything else.
/api/save-project's fallback parser could silently "succeed" while writing garbage filenames and dropping every real file, when a model wrapped its file markers in markdown heading/bold decoration instead of the exact instructed format — now parsed correctly, with fallback pattern ordering fixed so a loose heuristic can't out-run a precise one.- Docker installs baked a personal
config.json(save-dir path + a per-install calendar-feed secret) into the image, and only the Ollama service — not the app's own port — was bound to localhost-only indocker-compose.yml. Both fixed. - Update check/apply showed a misleading "could not reach GitHub" on installs with no
.gitdirectory (e.g. Docker); it now says so plainly instead. - Generated apps now carry a small "Powered by LLM CODER" credit — requested in the generation prompt, and always guaranteed in the project's README regardless of whether the model complied.
install.sh now detects GPU vendor/VRAM and system RAM on whatever machine it's run on and recommends/pulls a model sized to fit, instead of assuming a specific GPU. The RDNA4 ROCm gfx-version workaround only applies when that exact hardware is detected. docker-compose.yml no longer requires an Nvidia GPU to boot. An optional systemd --user service keeps the app (and Routines) running without a terminal open.
IMAP/SMTP via app passwords (Gmail, Outlook/Office 365, Yahoo, iCloud, Fastmail, Zoho, AOL, GMX, or custom). The AI drafts 3 reply options for you to pick from — nothing sends automatically. Calendar events get extracted from emails into a pending-approval queue; approved events push to CalDAV (where supported) and always publish to a universal ICS "subscribe by URL" feed, so even Google Calendar (which requires OAuth, not app passwords) can pick them up. Contacts are harvested automatically from your inbox.
Skills are reusable playbooks the agent can call on mid-task (get_skill tool) — write them by hand or have the AI draft one from a pasted conversation/project (you review before it's saved). Routines let you describe a recurring task in plain English; the AI reflects back what it understood, you Agree/Change/Cancel, then adjust the proposed schedule and confirm. Runs on an in-process scheduler with pause/resume/run-now and history.
Indexes your project files via Ollama embeddings (ollama pull nomic-embed-text) and searches them by meaning, not just filename — plus a semantic_search agent tool.
Desktop notifications (via notify-send) for finished routines, new pending calendar events, and new mail. One-click backup/restore (zip export of conversations/skills/routines/calendar data, account passwords redacted). Assistant code blocks get an Edit button next to Copy that opens a CodeMirror panel — apply changes back into the conversation.
The system prompt enforces zero restrictions. The model will never say:
- "I cannot" / "I can't" / "I'm not able to"
- "I'm sorry, but" / "As an AI, I cannot"
- "For legal/ethical/safety reasons"
It is free to assist with any request, any topic, any task without limitation.
- 💬 Chat — interactive conversation with local models (uncensored), with an in-chat code editor
- 🔍 App Analyzer — research and plan app development, or Clone Mode for full feature-parity replication of a real app (optionally grounded in a decompiled
.apk) - ⚡ Project Generator — describe an app, get a full project scaffold; Load a previously saved one back in
- ▶ Run Code — execute Python/JavaScript in a sandbox, or run a generated project live with an embedded browser preview
- 📱 Build Android APK — export a generated project as a real installable
.apkvia EAS Build - 🤖 Agent — autonomous multi-tool agent
- 📁 File Browser — manage project files
- 🌐 Search — web search (DuckDuckGo) plus semantic search over your own files
- 🎨 Image Generation — generate images via vision models
- 📦 Model Library — browse, download, and manage Ollama models
- 📧 Email — read, AI-drafted replies, never auto-sent
- 📅 Calendar — CalDAV + universal ICS feed, approval-queued event extraction
- 🧠 Skills — reusable, agent-invokable playbooks
- ⏰ Routines — scheduled AI tasks with a conversational setup flow
Linux / macOS:
./install.sh # detects your hardware, sets up Ollama + a fitting model, installs deps
./launch.sh # or install the auto-start service from install.sh (systemd on Linux, launchd on macOS) for persistent RoutinesWindows:
Download and run AI-Copper-Maker-Setup.exe (built from windows-installer/, see below) — it installs to your user profile (no admin rights needed), adds Start Menu/Desktop shortcuts, and offers to run first-time setup (installs Ollama, downloads a model, sets up the Python environment) at the end.
Or run the scripts directly from a checkout:
.\install-windows.ps1 # detects your hardware, sets up Ollama + a fitting model, installs deps
.\launch-windows.ps1 # or register the logon Scheduled Task from install-windows.ps1 for persistent RoutinesThen open http://localhost:8081
Select a model from the sidebar (e.g. qwen2.5-coder-32b-uncensored or mistral-24b-uncensored — or whatever the installer recommended for your hardware). For semantic search, also pull an embedding model: ollama pull nomic-embed-text.
windows-installer/build.sh # needs makensis (`brew install makensis`); outputs windows-installer/dist/AI-Copper-Maker-Setup.exeThe build stages only git-tracked files plus the Windows launcher scripts, so no local config, account credentials, or signing keystore ever end up in the package.
- Backend — FastAPI + Python 3, port 8081
- Frontend — single-page HTML/JS, served by FastAPI
- LLM — Ollama with local models (chat, embeddings)
- Code Execution — sandboxed temp directories, 30s timeout
- Web Search — DuckDuckGo HTML scraper (no API key needed)
- Image Gen — Ollama vision models (llava, minicpm-v, moondream)
- Email/Calendar — stdlib
imaplib/smtplib+caldav/icalendar - Scheduling — APScheduler (in-process; needs the backend running to fire)
- Code Editor — CodeMirror 6 (ESM, no build step)
- Python 3.11+
- Ollama (with your preferred models;
nomic-embed-textfor semantic search)