Quick Start: Domain Specialists¶
You're an expert in your field (research, medicine, law, finance) but not a software engineer. G6 helps you build more reliable applications that leverage your domain expertise through self-training, safety gates, tests, and optional formal verification backends.
How you reach G6¶
G6 is a hosted service. You use it from an MCP client on your own machine, which sends tool calls to our server. There is nothing to install and nothing to launch.
| Interface | Best for | How |
|---|---|---|
| MCP (the only customer interface) | Describing goals in natural language from inside your AI assistant | Setup guide |
claude mcp add g6 --transport sse \
--url https://g6solver.com/mcp/sse \
--header "Authorization: Bearer YOUR_API_KEY"
Any MCP client that supports SSE can connect the same way.
The GUI and TUI are not available
Earlier versions of this page recommended a browser dashboard (g6 --gui) and a terminal interface (g6 --tui) as the primary way in. Neither is part of the hosted product. They exist in the codebase and are documented under Bases for reference, but there is no desktop application to install, so there is no way to launch them. Everything below is done by asking your MCP client in natural language.
Your first session¶
1. Connect your client¶
Run the claude mcp add command above with the API key from your account dashboard. Then ask your assistant to use G6.
2. Describe your objective in domain terms¶
"Build a clinical decision support tool that recommends medication dosages based on patient weight, age, and renal function"
G6 decomposes this into verifiable subtasks and selects appropriate components. Every result carries a completion state — verified, qualified-draft, or blocked-escalated — so you can tell what has actually been checked from what merely looks finished.
3. Correct domain errors as they appear¶
When G6 proposes a plan that is wrong in a way only a specialist would catch, say so:
"The dosage calculation for paediatric patients should use the Schwartz equation, not Cockcroft-Gault. Correct this and remember it."
G6's self-training loop captures the correction as a stored theory and applies it next time.
4. Approvals¶
Where an action needs human judgement, G6 pauses and explains what it is about to do and why it is asking. You approve or redirect in the conversation.
What you get¶
- Self-training loop — G6 learns from your corrections and improves over time (T1 to T2 advancement)
- Formal verification — mathematical proofs that your application behaves correctly, not just test coverage
- Progressive disclosure — start simple, add complexity only when you need it
The T0-T3 framework¶
G6 classifies systems by their learning capability:
| Level | What it means | Your experience |
|---|---|---|
| T0 | No learning | Static tool use, no adaptation |
| T1 | Harness engineering | You fix problems once, G6 remembers |
| T2 | Continuous improvement | G6 learns from corrections automatically |
| T3 | Theory building | G6 discovers domain rules on its own |
Most domain specialists start at T1 (correcting G6) and advance to T2 (automated learning from corrections) within their first few sessions.
Next steps¶
- Setup guide — connecting your MCP client
- How Learning Works — explanation of the self-training loop
- Formal Methods — verification capabilities
- Safety & Alignment — how G6 ensures safe operations