Skip to content

TOOLS.md

Latest commit

 

History

History
61 lines (44 loc) · 3.45 KB

TOOLS.md

File metadata and controls

61 lines (44 loc) · 3.45 KB

FlowAI Supported Tools

Set master.tool and roles.<role-id>.tool in .flowai/config.json to one of these tool ids (each implemented as a plugin in src/tools/<name>.sh):

tool id Product Where it runs FlowAI behaviour
gemini Google Gemini CLI Terminal Full CLI session in the tmux pane; non-interactive phases add -m <model> -y.
claude Claude Code Terminal Full CLI in the pane; pipeline phases use --model and -p; --dangerously-skip-permissions when auto_approve: true. Optional MCP config via --mcp-config if .flowai/mcp.json exists.
cursor Cursor Terminal + Cursor app FlowAI prints the enriched prompt in the terminal for manual paste into Composer/Agent. Cursor routes internally to many providers; see Cursor models.
copilot GitHub Copilot Chat Terminal + Copilot Chat FlowAI prints the prompt for paste into Copilot Chat (no headless CLI available). Model routing is managed by GitHub.

Adding a new tool: create src/tools/<name>.sh defining three functions, then add the tools.<name> entry to models-catalog.json. No other files need to change.

Function Purpose
flowai_tool_<name>_print_models() Used by flowai models list
flowai_tool_<name>_run(model, auto_approve, run_interactive, sys_prompt) Phase dispatcher — runs full AI session
flowai_tool_<name>_run_oneshot(model, prompt_file) Non-interactive single-prompt for knowledge graph semantic extraction

Tools without a headless CLI (Cursor, Copilot) should return empty JSON from _run_oneshot:

printf '{"nodes":[],"edges":[],"insights":[]}'

Models and .flowai/config.json

Bundled catalog

FlowAI ships models-catalog.json at the repository root. For each tool it lists valid model ids, a default_id, optional notes, and links to vendor docs. Inspect with:

flowai models list          # all tools
flowai models list gemini
flowai models list claude
flowai models list cursor
flowai models list copilot

flowai init copies each tool's default_id from the catalog into the project config. After editing, run flowai validate. flowai start also validates on launch.

Config keys

Key Purpose
tool_defaults.<tool>.model Per-tool model override (any tool, new generic format)
default_model Gemini model override (legacy; still respected)
claude_default_model Claude model override (legacy; still respected)
master.model / roles.<id>.model Per-phase model (must be in catalog for gemini/claude)

At runtime, model ids not found in the catalog are replaced with default_id and a warning is logged. Use FLOWAI_ALLOW_UNKNOWN_MODEL=1 to pass any id through unchecked.

OpenAI-style ids (e.g. gpt-4o) with tool: "claude" are rejected and replaced with the Claude default.

Vendor references