Skip to content

Your First 30 Minutes with G6

This tutorial walks you through connecting G6, exploring its capabilities, running your first goal, and seeing the self-training loop in action.

Prerequisites

  • G6 installed — see Setup
  • Claude Code installed and working (for the MCP steps)

Minutes 0-3: Connect

Add the G6 MCP server to Claude Code:

claude mcp add g6 --transport sse \n  --url https://g6solver.com/mcp/sse \n  --header "Authorization: Bearer YOUR_API_KEY"

Verify the connection by asking Claude:

"Check the G6 system status"

You should see a response showing the system is healthy with 270 registry components available.

Minutes 3-8: Discover

G6 has built-in discovery tools. Try these natural language queries:

"What G6 tools are available for data analysis?"

This invokes nav_recommend and returns relevant components like adapt_pandas, adapt_sklearn, and align_evals.

"Give me an overview of the G6 system"

This invokes guide_system_overview and explains the architecture, component model, and available intelligence classes.

"Find G6 tools that can verify safety properties"

This shows the formal methods and CSF components — the tools G6 uses for formal property checks (proven where a named backend is installed and the property is formally specified) and for pre-execution risk bounds (CSF: policy priors, not guarantees).

Minutes 8-13: Your first goal

Let's decompose a simple goal. Ask Claude:

"Use G6 to decompose this goal: Research the current state of quantum error correction and summarise the key approaches in 300 words"

G6 will:

  1. Break the goal into subtasks (search, filter, synthesise)
  2. Identify which components to use (ctx_search, ctx_rag, LLM summarisation)
  3. Estimate resource requirements
  4. Return a SearchTree showing the decomposition

You'll see the goal broken into manageable steps with resource bounds for each.

Minutes 13-18: Run a pipeline

G6 has pre-built pipeline templates. Try:

"Run the G6 research pipeline on: advances in protein structure prediction since AlphaFold 3"

This executes the research template pipeline which chains:

  1. Search — finds relevant sources via ctx_search
  2. Retrieve — fetches and parses documents via ctx_rag
  3. Synthesise — produces a structured summary with citations

The result includes the summary, sources used, and confidence scores.

Minutes 18-23: Add safety

Now let's see the safety framework in action. Ask Claude:

"Use G6 to verify this claim using formal methods: If all inputs are positive and the function multiplies them pairwise, then all outputs are positive"

G6 will:

  1. Formalise the claim into propositional logic
  2. Run the built-in DPLL solver (no external dependencies needed)
  3. Return SAT/UNSAT with a proof or counterexample

This is the same formal verification pipeline CSF uses to check safety properties before code execution — proven when the named backend returns verified=True, otherwise recorded as auditable evidence, not proof.

Minutes 23-28: Self-training preview

The self-training loop is G6's core innovation. Here's how it works conceptually:

"Use G6 to explain how the self-training loop works"

The 7-step cycle:

  1. Audit — assess current harness maturity
  2. Observe — gather execution traces
  3. Orient — classify failure modes
  4. Diagnose — identify root causes
  5. Decide — choose improvement strategy
  6. Improve — modify prompts, tools, or logic
  7. Record — update competence model

Each iteration improves the harness without retraining the underlying model. This is what enables G6 to achieve 58-77% accuracy improvements on validated benchmarks.

Minutes 28-30: What's next

You've now seen G6's four core capabilities:

Capability What you tried
Discovery nav_recommend, guide_system_overview
Goal decomposition Breaking a research task into subtasks
Formal verification Proving a mathematical property
Pipeline execution Running the research template

Next steps

Explore by persona