Documentation
Quick Start
Grimoire turns a plain-English description into a working, documented codebase pushed to a private GitHub repo on your account. Pick whichever surface you already live in:
# Browser — zero install (keys stay in your browser)
open conjure.html on this site → paste a GitHub or GitLab token + a Gemini, Claude, Groq or OpenRouter key → CONJURE
# Discord — in any server the bot has joined
/build a recipe app with search
/model # switch provider / model
# Self-host the engine (CLI, bot, everything)
git clone https://github.com/Arnav1771/turmux_builder
pip install -r requirements.txt
python main.py # .env holds your keysEvery Surface
| Surface | How |
|---|---|
| Web (this site) | Conjure App — 100% client-side, BYO keys |
| Claude Agent Skill | say “Grimoire, build me a…” in Claude Code / Claude.ai |
| MCP server | claude mcp add grimoire -- grimoire-mcp |
| CLI / Termux | grimoire build "…" — laptop or phone |
| Discord bot | /build · /model |
The Two-Pass Engine
Pass 1 — Plan. The model returns structured JSON: repo name, tech stack, and a full file tree with a purpose for each file.
Pass 2 — Generate. Each file is written in its own call with the whole plan as context, so imports and structure line up. The result is committed file-by-file to a new private repo through the GitHub API — no local git required.
Bring Your Own Keys
There is no backend and no account. In the browser, keys persist only in localStorage (opt-out checkbox) and are sent directly to your provider and GitHub over HTTPS. Self-hosting reads keys from .env.
IMP_DOCS
Every conjured repo ships a standardized IMP_DOCS/ folder — HANDOFF, PROMPT_TRAIL and friends — plus HOW_TO_RUN.md and a CI validate-build.yml, so a teammate can pick the project up cold.
Providers & Models
| Provider | Examples | In-browser |
|---|---|---|
| Google Gemini | gemini-3.8-flash · gemini-3.1-pro-preview | ✓ |
| Anthropic Claude | claude-opus-5-5 · claude-sonnet-5 · claude-haiku-4-5-20251001 | ✓ |
| Groq | openai/gpt-oss-120b · qwen/qwen3.8-27b (preview) | ✓ |
| OpenRouter | nvidia/nemotron-3-super-120b-a12b · openai/gpt-6-sol · :free variants | ✓ |
| NVIDIA (direct API) | Nemotron via integrate.api.nvidia.com | CLI / MCP / Discord — no CORS; in the browser use OpenRouter |
| GitHub Models | — | Retired by GitHub on 30 July 2026 |
Every provider needs your own key. A provider can only run in the browser if its API grants CORS to a web page: Gemini, Anthropic, Groq and OpenRouter do, NVIDIA's own API does not, which is why Nemotron comes through OpenRouter here. OpenRouter's :free models are rate-limited to 20 requests a minute and 50 a day until the account has bought $10 of credits.