Embed badge
Show
Style
[](https://versuz.dev/claude-md/willnigri-agentic-tool-optimization)A CLAUDE.md is just a markdown file at the root of your repo. Copy the content below into your own project's CLAUDE.md to give your agent the same context.
npx versuz@latest install willnigri-agentic-tool-optimization --kind=claude-mdcurl -o CLAUDE.md https://raw.githubusercontent.com/WillNigri/Agentic-Tool-Optimization/HEAD/CLAUDE.md# ATO (Open Source) The GUI for creating, managing, and observing AI agents — across Claude Code, Codex / OpenAI Agents SDK, Gemini CLI / ADK, OpenClaw, Hermes, and Ollama. MIT licensed. Local-first. Positioning as of v1.3.0 (The GUI Pivot): "the place where you create an agent, not just manage one." Three audiences served by one app: non-technical first-timers (guided chat wizard), power users (Quick form + ⌘K + terminal), teams (Pro/Team tier in `ato-cloud`). See `PRD.md` and `ROADMAP.md` for full positioning. ## Structure ``` apps/desktop/ # Tauri 2.x desktop app (Rust + React) packages/core/ # Shared types, token utils, config paths (no I/O) packages/db/ # Database abstraction (SQLite for desktop) services/mcp-server/ # Standalone MCP server for AI coding tools (stdio, 8 tools) ``` ## Commands - `npm run dev:desktop` — Start Tauri desktop app in dev mode - `npm run dev:mcp` — Start MCP server in dev mode - `npx tauri build` — Build desktop app for distribution - `npm run build` — Build all packages ## Desktop App Tauri 2.x with: - **Rust backend**: SQLite (rusqlite), multi-runtime CLI dispatch, recursive skill scanning, context estimation, file watcher (notify crate) - **React frontend**: Vite + TailwindCSS + Recharts + Zustand + TanStack React Query - **i18n**: English, Portuguese, Spanish (react-i18next) - **Theme**: Dark (#0a0a0f) + cyan/mint (#00FFB2) accent - **Offline-first**: Works without internet, all data in local SQLite - **withGlobalTauri**: Enabled in tauri.conf.json for Tauri API access ## Multi-Agent Runtime (Two-Way) Supported runtimes: Claude, Codex, OpenClaw (SSH), Hermes. **Outbound**: `promptAgent(runtime, prompt, config?)` — unified dispatch **Inbound**: `queryAgentStatus(runtime)` — deep health check (version, auth, connectivity) **Logging**: All executions auto-logged to `~/.ato/agent-logs.jsonl` **Detection**: Searches common paths + npx cache + user's shell PATH + manual override ## Skill Directories Per Runtime - Claude: `~/.claude/skills/`, `.claude/skills/`, `/etc/claude/skills/` - Codex: `~/.codex/skills/`, `.agents/skills/`, `.codex/skills/` ($CODEX_HOME) - OpenClaw: `~/.openclaw/skills/`, workspace/skills/, AGENTS.md, SOUL.md, TOOLS.md ($OPENCLAW_HOME) - Hermes: `~/.hermes/skills/` (with category subdirs), SOUL.md, memories/ ## MCP Server Tools - `get_context_usage` — Context window breakdown - `list_skills` / `toggle_skill` — Manage skills - `get_usage_stats` — Token/cost analytics - `get_mcp_status` — MCP server health - `get_runtime_status` — Health check for any runtime - `get_all_runtime_statuses` — All runtimes at once - `get_agent_logs` — Execution logs (filterable by runtime) ## Key Implementation Details - **Project paths**: `project_root()` walks up from CWD to find `.git` or `.claude/` (Tauri CWD is apps/desktop/) - **Skills are on-demand**: NOT counted in context total (loaded only when triggered) - **Automation flows**: Auto-detected from `## Step N:` and `## Phase N:` headers in SKILL.md - **Approval dialog**: When agent response contains SKILL.md content, shows approval UI - **Setup wizard**: First-launch onboarding, persisted to localStorage ## Information Architecture (v1.3.0+) The desktop app sidebar has **6 top-level sections** (collapsed from 24 in v1.2.x). Every screen lives under one of these: 1. **Home** — landing; Create Agent CTA, recent agents, recent runs, alerts. 2. **Agents** — list, detail (Config / Skills / MCPs / Permissions), + New Agent wizard. 3. **Skills & MCPs** — Skills tab, MCPs tab (with install registry), Marketplace tab. 4. **Runs** — Live (active sessions) / History (logs) / Schedules (cron) / Automations (workflows) / Hooks. 5. **Insights** — Health, Analytics, Context, Audit Log. 6. **Settings** — Runtimes (merged), Models, API Keys, Secrets, Env, Cloud (auth+sync+teams+notifications), Projects. **Persistent surfaces**: Command palette (⌘K), Chat/Terminal pane (bottom, expandable), Project switcher (sidebar top). ## Create Agent flow (v1.3.0+) Two paths share the same end-state. Guided (chat) is default for non-technical users; Quick (form) is default for power users. **File-writing contract per runtime** (all writes go through hash check + auto-backup + audit log): | Runtime | Path | |---|---| | Claude Code | `~/.claude/agents/<slug>.md` (global) or `<project>/.claude/agents/<slug>.md` (project) | | Codex / OpenAI Agents SDK | `~/.codex/agents/<slug>/` directory + `AGENTS.md` | | Gemini CLI / ADK | `<project>/.gemini/agents/<slug>.yaml` (root_agent.yaml entry) | | OpenClaw | `~/.openclaw/agents/<slug>/SOUL.md` + `TOOLS.md` | | Hermes | `~/.hermes/agents/<slug>/` | **Auth resolution order** for the wizard's "AI suggests a stack" step (subscriptions OR keys, both first-class): 1. Active CLI subscription — shell out (`claude --print`, `codex --print`, `gemini -p`) 2. Active API key in `LlmApiKeys` — direct provider SDK 3. Local Ollama on `localhost:11434` 4. Pro-tier `/agent-suggest` on `ato-cloud` (fallback) Drafts persisted to `~/.ato/agent-drafts/<id>.json` so users can leave and resume. ## Embedded Terminal (v1.3.0+) Replaces chat-only PromptBar with an expandable bottom pane, two modes: - **Chat mode** — same as today's PromptBar (`promptAgent()`). - **Terminal mode** — full xterm shell scoped to active project CWD. **Stack**: - Frontend: `xterm.js` + `@xterm/addon-fit` + `@xterm/addon-web-links` - Backend: `portable-pty` Rust crate, spawned via Tauri command, default shell `$SHELL` (Unix) / `pwsh.exe` (Windows) - PATH from login-shell spawn (existing pattern). - Persistent across navigation; collapsible; resizable. ## Open Source vs Closed Source **Open source (this repo)**: Skills manager + marketplace, multi-agent runtime, subagents, **Create Agent wizard (v1.3.0)**, **MCP install UI (v1.3.0)**, **embedded terminal (v1.3.0)**, automation builder, cron scheduling, context visualizer, hooks, ATO MCP server (8 tools), config editor, command palette, i18n. **Closed source (`ato-cloud`, paid)**: Real-time cron monitoring, silent failure detection, push notifications, usage analytics (cloud), cloud sync, team workspaces, hosted `/agent-suggest` fallback, MCP registry mirror, hosted templates library, configuration backup vault, hosted terminal sessions (Team). ## Security - Desktop is local-first. No network calls unless sync explicitly enabled. - Use parameterized SQL queries only. - SSH for OpenClaw uses key-based auth (paths only, no key contents stored). - Validate all inputs with zod schemas. - Tauri apps get user's full PATH via login shell spawn.