REPL¶

Not part of the hosted product

G6 is hosted: you reach it over MCP by pointing Claude Code (or any MCP client) at the server. There is no desktop app to install and no self-hosted deployment. This page documents a surface that exists in the repository but is not part of the hosted product today — it is kept as engineering reference, not as instructions for customers.

The REPL is an interactive prompt-toolkit shell with agentic LLM chat. It is the default mode when you run g6 with no subcommand.

Overview¶

# Launch REPL (default)
g6

# Force REPL mode
G6_CLI_MODE=click g6

The REPL combines two interaction modes:

CLI commands — type any CLI command directly (e.g. invoke core health, registry list)
Natural language — queries that don't match a command are routed to an LLM for agentic chat with tool-calling

Workspaces¶

Each REPL session runs inside a workspace — a directory under ~/.g6/g6_workspaces/<name>/ that persists:

Chat history (chat_history.jsonl)
Session state
Run artifacts
Git-backed versioning via WorkspaceManagerBlock

On startup, the REPL offers to resume the last workspace or create a new one.

LLM agentic chat¶

Natural language queries are handled by an LLM agent with access to 10 tools. The agent can chain up to 8 tool calls per query to research and answer your question.

Backends¶

Backend	Provider	Model	Selection
Ollama	Local	`gemma3:12b` (configurable)	Default — no API key needed
OpenRouter	Cloud	`google/gemini-2.5-flash`	Requires `OPENROUTER_API_KEY`
Claude Code	Local subprocess	Inherited from host	Set `G6_REPL_BACKEND=claude`

The REPL auto-selects Ollama if available, falling back to OpenRouter if an API key is set.

Tool-calling¶

When the LLM backend supports function calling, the agent has access to these tools:

Tool	Description
`read_file`	Read a file from the workspace (max 8,000 chars)
`list_dir`	List files and directories
`fetch_url`	HTTP GET with HTML stripping (max 12,000 chars)
`list_components`	List all G6 components with op counts
`nav_search`	Search components by keyword
`guide_ask`	Ask the G6 guide a question
`invoke_component`	Invoke a component operation
`decompose_goal`	Break a goal into subtasks
`run_pipeline`	Run a component pipeline
`calculator`	Evaluate mathematical expressions

The agent executes up to 8 tool-calling steps per query. A terminal spinner shows progress during LLM calls.

Cross-session memory¶

The REPL maintains two memory layers for context across sessions:

Layer	Technology	Purpose
Working memory	ACE (TF-IDF)	Fast recall of recent interactions
Long-term memory	Cognee (knowledge graph)	Entity and relationship extraction

Memory is per-workspace and loaded/saved automatically on workspace open/close.

Built-in commands¶

Command	Description
`help`	Show available commands
`clear`	Clear conversation history
`workspace`	Show current workspace info
`tui`	Launch the TUI dashboard
`quit` / `exit`	Exit the REPL

All CLI commands are also available directly in the REPL (e.g. registry list, hitl list, session status).

Configuration¶

Variable	Default	Description
`G6_CLI_MODE`	`both`	`both` / `tui` / `click` — controls default launch mode
`G6_REPL_BACKEND`	(auto)	Force backend: `claude`, `openrouter`, `ollama`
`G6_ASCII_MODE`	(unset)	Set to `1` for ASCII-only rendering (no Unicode)
`SOLVER__DEFAULT_LLM_BACKEND`	`ollama`	Default LLM backend for all G6 operations
`SOLVER__DEFAULT_OLLAMA_MODEL`	`gemma3:12b`	Ollama model to use
`OPENROUTER_API_KEY`	(unset)	Required for OpenRouter backend

Source files¶

File	Purpose
`bases/mvp/cli/g6_cli.py`	REPL loop, workspace management, LLM dispatch
`bases/mvp/cli/repl_tools.py`	10 tool schemas and executors
`bases/mvp/cli/repl_skin.py`	Prompt styling and colour themes
`bases/mvp/cli/memory_manager.py`	ACE + Cognee memory integration
`bases/mvp/cli/mcp_client.py`	MCP client for tool-calling