Frequently Asked Questions¶

General¶

What is G6?¶

G6 is a composable AI framework that exposes 270 AI components as MCP tools for Claude Code. It adds formal reasoning, ML pipelines, safety verification, self-training, and persistent memory to your AI assistant — invokable through natural language.

What problems does G6 solve?¶

G6 addresses the gap between what LLMs can theoretically do and what they reliably deliver in production:

• Reliability: Self-training loops that improve harness accuracy from baseline to 80%+ without model retraining
• Safety: Computational Safety Framework (CSF) with configurable pre-execution risk bounds -- policy priors, not formal proofs
• Cost: Hyperdistillation transfers frontier-model intelligence into cheaper local models
• Trust: Formal verification (17 solver backends) can prove properties before execution when the relevant backend is installed and the property is formally specified; otherwise G6 records the check as auditable evidence, not proof

Do I need ML expertise to use G6?¶

No. G6 is designed for three audiences:

• Business operators: Use the GUI to create goals and approve results
• Domain workers: Use MCP tools through natural language in Claude Code
• Engineers: Build custom components and pipelines for production deployment

Setup¶

What are the system requirements?¶

• Windows 10+, macOS 12+, or Ubuntu 22.04+
• 4 GB RAM (8 GB for ML workloads)
• Claude Code with MCP support (for the optional Claude Code integration)

How do I install G6?¶

There is nothing to install. G6 runs on G6Solver's servers; you connect your own MCP client to it. Create an account, generate an API key, and follow the connect instructions on the setup page, which carries the current transport and URL. The command has the shape:

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

See Setup for detailed instructions.

Does G6 work offline?¶

No. G6 runs on G6Solver's servers, so every tool call needs network access to reach the hosted MCP endpoint. There is no local engine to fall back on. What runs on your machine is your own agent (Claude Code, your script); G6 itself does not.

Which LLM providers are supported?¶

• Ollama — local models, no cost, full privacy (recommended for development)
• OpenRouter — access to 100+ models including Claude, GPT-4, Gemini, Llama
• Claude Code — headless mode for agentic tasks (uses your existing subscription)

Pricing¶

What's included in the free trial?¶

30 days of full access to the Builder tier (all 270 registry components, 500 calls/min). No credit card required. After trial expiry, you drop to the free tier (30 core tools, 10 calls/min).

What's the difference between Researcher and Builder?¶

Are there usage limits?¶

Rate limits vary by tier (10/100/500 calls per minute). There are no hard caps on monthly usage for subscription tiers. PAYG customers pay per call at 1.5x the subscription rate.

How much does a typical task cost?¶

Costs depend on which LLM backend is used. Ollama (local) = $0. OpenRouter varies by model.

Technical¶

How does G6 differ from LangChain or CrewAI?¶

G6 sits at a higher level: it orchestrates the construction of harnesses rather than being a harness itself.

Can I use G6 with non-Claude models?¶

Yes. G6 supports Ollama (any local model), OpenRouter (100+ models), and direct API calls. The MCP integration requires Claude Code, but the REST API works with any client.

How are safety guarantees enforced?¶

The Computational Safety Framework (CSF) provides:

1. Pre-execution bounds: CSF hazard probabilities calculated before any action
2. Guardrails: Configurable thresholds that halt execution if violated
3. Formal verification: Properties can be proven via Z3/DPLL/Lean when the relevant backend is installed and the property is formally specified; otherwise the check is recorded as auditable evidence, not proof
4. Breakpoints: Human-in-the-loop approval gates for high-risk operations
5. Rollback: Git-based workspace isolation with automatic rollback on failure
6. Global kill switch: One-button emergency stop across all interfaces (MCP/REST/GUI/TUI)
7. Stable snapshots: Git-tag-based versioning with automatic snapshots after tests pass, manual marking, and instant restore

What is the kill switch and how do I use it?¶

The global kill switch is a system-wide emergency stop. When activated, all autonomous operations halt immediately — orchestrators, coding agents, SDLC workflows, and recursive architect branches.

• MCP: emergency_stop / emergency_resume / emergency_status tools
• REST: POST /emergency/stop / POST /emergency/resume / GET /emergency/status
• TUI: Press Ctrl+X for emergency stop. The status bar shows "HALTED" in red when active.

The kill switch uses a file sentinel (~/.g6/KILL_SWITCH). Deleting the file manually also deactivates it.

What are stable snapshots?¶

Stable snapshots mark known-good states of your workspace using git tags. Old snapshots are never deleted — they're append-only.

Three types of snapshots:

1. Manual: You mark the current state as stable via MCP (create_stable_snapshot), REST (POST /emergency/snapshots), or TUI
2. Automatic (test pass): Created automatically when the test suite passes
3. Pre-evolution: Created automatically before any T3 self-modification event

To restore: use restore_stable_snapshot (MCP) or POST /emergency/snapshots/{id}/restore (REST). This performs a git reset --hard to the tagged commit.

Can I build custom components?¶

Yes. See the Custom Component tutorial. A component requires:

1. Input/Output Pydantic models
2. An AIBlock subclass with an infer() method
3. Tests in tests/mvp/<component_name>/

Troubleshooting¶

Tool calls are failing — what should I check?¶

1. Is the MCP server running? Check with system_status tool
2. Correct tier? Some tools require Researcher or Builder tier
3. Rate limited? Check your calls/minute against your tier limit
4. Backend available? If using Ollama, verify it's running (ollama list)

I'm hitting rate limits — what can I do?¶

• Upgrade your tier (Free: 10/min, Researcher: 100/min, Builder: 500/min)
• Batch operations using the pipeline tools instead of individual calls
• Use local backends (Ollama) which have no rate limits

The MCP server won't start¶

G6's MCP server runs on G6Solver's servers — there is nothing to start locally. If your client cannot reach it, check:

• Authentication: your API key must be sent as Authorization: Bearer <key>. Regenerate the key from your dashboard if in doubt.
• Transport and URL: the client must use the SSE transport and the URL shown on the setup page. A stale connect string is the most common cause.
• Subscription state: an expired trial drops you to the free tier, which exposes fewer tools; a tool that "disappeared" is usually a tier change.