Documentation

Getting Started

Create an account, generate an API key, and connect Claude Code to the hosted G6Solver MCP endpoint in under 5 minutes.

Setup

G6Solver is a hosted service. You create an account, generate an API key, and point your MCP client at our endpoint. There is no G6 software to install on your machine.

Create your account

Generate an API key

Go to Account → API Keys and press Generate. Copy the key immediately — it is shown once and never again.

The key authenticates every request. Anyone holding it can use your account; revoke it on the same page if it leaks.

Connect Claude Code

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

Any MCP client with SSE support works the same way. Then ask your client to call g6_first_call_help_tool for the recommended first session. The guided setup walks through the same four steps with your progress tracked.

Choose Your Interface

G6 has four interfaces. Only one is available to customers today:

Available now: the MCP plugin, over our hosted endpoint. The Browser GUI, Terminal TUI and Interactive REPL all run from a local g6 executable, which is not currently distributed — they are listed here for completeness, not as options you can pick.

Browser GUI

Visual dashboard with run management, HITL approvals, learning tools, and artifact inspection. Best for business operators and domain specialists.

g6 --gui

6 views: Threads, Projects, Library, Jobs, HITL, Settings

Terminal TUI

17-tab terminal interface with chat, tasks, observability, job agents, and flywheel accuracy tracking. Best for terminal-comfortable users.

g6 --tui

17 tabs: Chat, Progress, Tasks, Observability, Flywheel, Files, Components, Jobs, TODO / Review, Settings, Safety, Gate, Reliability, T3, Autofix, Results, Demo

MCP Plugin

Add G6 as a plugin inside Claude Code, Cursor, or any MCP-compatible agent. Describe goals in natural language and let the agent call G6 tools.

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

Tier-gated tools across 15 categories, executing on G6’s servers and returned to your coding agent

Interactive REPL

Command-line shell with agentic LLM chat. Type CLI commands or ask questions in natural language. Cross-session memory remembers context.

g6

10 agentic tools, OpenRouter LLM routing plus a Claude Code coding-agent backend, persistent workspaces. (An Ollama local-GPU backend exists in the codebase but is not available on the hosted service until a desktop application ships.)

Which interface should I use?

I am a...	Start with	Why
Business Operator	g6 --gui	Point-and-click run management, visual HITL approvals
Domain Worker	g6 --gui or g6 --tui	GUI for browsing job agents, TUI Jobs tab for terminal users
Software Engineer	MCP or g6 (REPL)	MCP integrates with your coding agent; REPL for quick queries
Domain Specialist	g6 --gui	Learning dashboard for training, TUI Flywheel for accuracy tracking

Tool Access by Plan

Not all tools are available on every plan. Here’s what you get:

Free

Core + Guide + Invariants + Diagnostics

Researcher

+ Navigator, EvoSkill (read), Bug reports

Full

Builder — full tool access

All 15 categories

All plans install the same binary — your license tier gates which MCP tools are available. invoke_component is in the free tier, so any user can invoke any installed component directly. Trial accounts get the Researcher tool set for 30 days, then drop to Free. See MCP Tools for the full breakdown.

Natural Language Goals

Don’t want to write JSON? G6’s GoalWizard translates plain English into structured goals with appropriate resource bounds, guardrails, and checkpoints.

# Just tell Claude what you want:
"Use G6 to analyse our Q4 sales data and forecast next quarter by region"

# GoalWizard automatically sets up:
# - Resource bounds (token limits, time caps)
# - Guardrails (budget constraints)
# - The right components (adapt_pandas, adapt_sklearn, align_evals)

Why G6?

G6 extends Claude Code with capabilities that no single LLM session can provide:

Persistent Learning (T1–T2)

Claude Code forgets everything between sessions. G6 remembers validated solutions and builds on them across sessions — one adaptive traverse beyond what any single LLM session can achieve. The SOAR case bank and distillation library persist to disk and are read back on the next run. The mastery tracker persists too, but nothing reads it back yet.

Formal Verification

Claude Code cannot verify its own outputs mathematically. G6 can route explicitly specified properties to SAT solvers, SMT checkers, and proof tools when those backends are installed; otherwise it records static checks and generated specs as auditable evidence.

Self-Improving Skills (T2–T3)

Claude Code treats every problem as new. G6’s evolution loop discovers better strategies and writes standalone algorithms — learning how to learn, not just what to learn. Cross-session learning is wired for two of the four stores: the SOAR case bank and the distillation library are written and read back. The mastery tracker and the Thompson-sampling meta-learner are written but not yet read, so they do not influence a later run.

Cost Optimisation

As the template library grows, recurring patterns are handled by cached algorithms instead of fresh LLM calls. The distillation pipeline is fully wired — the template library is in early development, so most tasks still use the full LLM reasoning path.

Configuration Reference

Environment Variables

Variable	Purpose	Required
OPENROUTER_API_KEY	LLM routing via OpenRouter	Yes
ANTHROPIC_API_KEY	Direct Claude API access	Optional
PANDAS_DB_PATH	Pandas MCP SQLite location (default: ~/.pandas_mcp/pandas.db)	No

Security Notice

Never commit .env to version control. Use environment variables or a secret manager in production.

Config Settings

These settings are configured in G6 Config and can be overridden via environment variables:

Setting	Default	Description
allowed_llm_base_urls	["https://openrouter.ai/api/v1"]	Permitted LLM API base URLs
max_llm_tokens_cap	128000	Maximum token limit per LLM request
default_llm_timeout_sec	60	Default timeout for LLM calls (seconds)

Your First Session

Once G6 is connected to Claude Code, all tools are available immediately. Try these examples:

Check System Health

Ask Claude to call the system_status tool to verify your connection.

# In Claude Code, simply ask:
"Check the G6 system status"

G6 returns:

G6 Hyperdistillation system: OPERATIONAL
Available tools: full tool access
Components: 200+
Version: 0.1.0

Decompose a Goal

# Ask Claude:
"Use G6 to decompose: Analyze this codebase and identify potential memory leaks"

G6 returns a structured subtask tree:

{
  "goal": "Analyze this codebase and identify potential memory leaks",
  "subtasks": [
    "Parse source files and build AST representation",
    "Identify allocation patterns without corresponding deallocation",
    "Check for circular references in object graphs",
    "Generate report with file locations and severity ratings"
  ],
  "ok": true
}

Run a Pipeline

# Ask Claude:
"Use the G6 run_pipeline tool to execute a safety verification pipeline"

G6 orchestrates the pipeline and returns step-by-step results:

{
  "ok": true,
  "result": [
    {"step": 1, "component": "meta_programming", "status": "ok"},
    {"step": 2, "component": "formal_methods", "status": "ok"},
    {"step": 3, "component": "align_csf", "status": "ok", "safety_score": 0.94}
  ]
}

List Available Components

# Ask Claude:
"List all available G6 components"

G6 returns all discoverable components:

{
  "components": [
    "core", "goal_engine", "formal_methods", "adapt_sklearn",
    "adapt_memory", "ctx_rag", "grounding", "agent_claude",
    "align_csf", "meta_programming", "evoskill",
    ... (200+ total)
  ]
}

5-Minute Recipe: Complete Workflow

1. system_status — Verify G6 is connected and see available tools
2. decompose_goal — Break your objective into verified subtasks
3. run_pipeline — Execute the subtasks as a composable pipeline
4. run_safety_pipeline — Check outputs with safety gates and available formal backends

As G6 learns your patterns, recurring problems are solved more efficiently. The SOAR case bank and distillation library persist across sessions, so the solver starts each run with accumulated knowledge. See benchmarks for current cost data.

End-to-End Walkthrough

Here's what a real conversation with G6 looks like in Claude Code:

Step 1 — You ask Claude:

"Check the G6 system status"

Claude calls system_status → G6 returns:

G6 Hyperdistillation system: OPERATIONAL
Available tools: full tool access

Step 2 — You ask Claude:

"Use G6 to find the best ML algorithm for my classification dataset"

Claude calls recommend_algo → G6 returns:

Task: classification
Top recommendation: RandomForest (score=0.92)
Reason: Balanced accuracy/speed for tabular data with 12 features

Step 3 — You ask Claude:

"Now use G6 to train and evaluate that model with run_automl"

Claude calls run_automl → G6 returns:

Best model: RandomForest
Score: 0.95 (accuracy)
Models tried: 12
Task: classification

Step 4 — You ask Claude:

"Which G6 component can help me verify these results formally?"

Claude calls guide_find_tool → G6 returns:

{
  "ok": true,
  "answer": "Use the formal_methods component for available checks, then inspect the proof status and artifact",
  "recommendations": ["formal_methods", "align_evals", "align_csf"]
}

REST API Quick Start

REST is not part of the hosted product today. The /api/ route is closed at the g6solver.com edge, so the calls below return 404 there. MCP is the surface to use; these are kept for anyone running the REST base themselves.

Health Check

curl https://g6solver.com/api/health

List Components

curl https://g6solver.com/api/components \
  -H "Authorization: Bearer YOUR_API_KEY"

Invoke a Component

curl -X POST https://g6solver.com/api/invoke/<component> \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"op": "...", "params": {}}'

See the API reference for the component list and each component’s operations.

Troubleshooting

401 Unauthorized

{"ok": false, "error": "Invalid or missing API key"}

Check that your client is sending Authorization: Bearer YOUR_API_KEY, and that the key has not been revoked on Account → API Keys. Keys are shown once at creation; if you lost it, generate a new one and revoke the old.

429 Rate Limited

{"ok": false, "error": "Rate limit exceeded. Retry after 12s"}

Rate limits are enforced per API key at your tier’s requests-per-minute ceiling. Wait for the retry period, or upgrade for a higher limit.

Tool Not Available on Your Tier

{"ok": false, "error": "Tool 'evolve_skills' requires Builder plan"}

Some tools are only available on higher tiers. See the MCP tool reference for tier requirements.