Charlotte
Decide which model does the work — and run it on hardware you own
Routing, not another chat window
Most teams paying for AI are paying premium per-token rates for work a smaller or local model would have handled just as well. Charlotte sits in front of the models and makes that call per task: it classifies the work, picks an engine, runs it, and returns the result through one surface.
The design point is that the cheap path is the default path. Local models on your own GPU take what they can take; a premium API is what you escalate to, not what you start with.
What this is today: an operator's desktop application and command-line tool that run on your own machine — not a hosted service. There is no Charlotte signup page. It is the control plane we run our own engineering on, and we are talking with a small number of teams about running it on theirs.
What runs today
Two live engines, honestly counted
The routing API registers exactly two adapters at startup — the Claude Code CLI and local Ollama. Cloud adapters activate only when a key is present. The registry reports what is wired, not what could be.
Seven local models
llama3.1:8b, llama3.2:3b, qwen2.5:7b, qwen2.5-coder:7b, qwen2.5:14b, deepseek-r1:14b and nomic-embed-text, under Ollama on a single 16GB consumer GPU.
A desktop console
Dispatch, Voice, Voice Lab, Reader, Models and Sigil tabs in one Windows application. The Sigil tab drives a live tenant end to end — login, sites, plugin toggles.
Voice synthesis on-box
Kokoro TTS with 54 voices, VoxCPM, and IndexTTS-2.5 zero-shot cloning against an 18-voice reference bank, through one serialized playback queue.
Documents to narrated scenes
The Reader pipeline turns a written document into story beats, per-beat voiced audio and scene environment data — an export driven end to end into a real Unreal project.
A Rust command line
route, dispatch, rag, vault, standards, an MCP server, and a sigil verb family that is a working HTTP client against the production Sigil API, acceptance-tested 2026-09-24.
Spending real money on AI APIs?
Charlotte is the routing layer we run our own engineering on. If you are paying premium rates for work a local model could absorb, we should talk.