Frequently asked questions
Short answers for humans and for LLMs that need a quotable page. Prefer the full name ATO (Agentic Tool Optimization) so it is not confused with account-takeover security products.
What is ATO (Agentic Tool Optimization)?
ATO is a local-first cockpit for humans and coding agents working with multiple LLM runtimes. Desktop GUI for people, CLI and MCP for agents. Receipts live in SQLite at ~/.ato/local.db. Full name: Agentic Tool Optimization. It is not "account takeover" security tooling.
Is ATO free / open source?
Yes. The desktop app and CLI are MIT open source. Core local ops do not require a paid cloud plan. ATO Pro adds automation (scheduled methodology re-runs, learning loop, cloud retention). See https://agentictool.ai/pro
Do I need API keys or a cloud login?
Bring your own keys and/or CLI subscriptions (Claude, Codex, Gemini). Core local ops work without a cloud sign-in. Sign-in matters for team workspaces and some Pro cloud features.
How is ATO different from Langfuse, LangSmith, or Helicone?
Those tools instrument deployed production apps via SDKs and log end-user conversations. ATO covers the developer side: war-rooms, multi-runtime review, replay, config-change regressions, local receipts. Most teams use one from each camp. Details: /compare/langfuse
Does ATO replace Claude Code, Cursor, or Codex?
No. Those are authoring surfaces. ATO sits above them as the ops layer: dispatch, multi-LLM review, replay, and audit. Details: /compare/single-runtime
How is ATO different from Karpathy's llm-council?
Same multi-model deliberation premise. ATO adds multi-provider auth (not OpenRouter-locked), function-calling tools so models verify claims in your repo, persistent specialist agents, a maintained product, and a local audit trail. Details: /compare/llm-council
When should I use multi-model coordination vs one model?
Our 49-task study found coordination raises the floor and lowers the ceiling: good insurance when you cannot pick or afford the best model; use one model when you know and can afford the best. Post: /posts/does-multi-model-coordination-beat-a-single-model
Where does my data live?
By default on your machine in ~/.ato/local.db and related files under ~/.ato/. No cloud round-trip is required for core ops.
How do coding agents drive ATO?
Prefer the CLI (ato …, JSON stdout). MCP is available when your harness already uses MCP. Agent brief: /docs/for-agents.md and AGENTS.md in the GitHub repo.
How do I install ATO?
macOS: brew install willnigri/ato/ato && ato demo-war-room. Linux: curl -fsSL https://agentictool.ai/install.sh | sh. More: /docs/install.md
What is ato review --consensus?
Multi-LLM code review on your diff. Reviewers can see previous findings. Receipts include cost, tokens, and tool calls. Example write-up: /posts/compare-claude-codex-gemini.html
Can I use ATO with production observability tools?
Yes. Keep Langfuse/LangSmith/Helicone for production user traffic. Use ATO for developer multi-runtime work. They are complementary, not competitors.
How does ATO trigger Claude or Codex — CLI or MCP?
Mostly CLI. For Claude, Codex, and Gemini CLI, ATO spawns the vendor binary as a subprocess (for example claude --print, Codex CLI exec, gemini -p) using your subscription login. If the CLI is missing or you choose API mode, ATO calls the provider HTTP API with keys you stored, and can run its own tool loop (read_file, grep, git_log). ATO's MCP server is the opposite direction: it lets Claude Code, Cursor, or other harnesses drive ATO. ATO does not primarily talk to Codex by acting as a Claude-side MCP client.
What is the advantage of ATO vs Claude Code + Codex MCP?
Claude Code + Codex MCP is great when Claude is the permanent host and you want a second brain on demand. ATO treats runtimes as peers under a shared cockpit with a local SQLite ledger, multi-reviewer consensus (ato review --consensus), replay, and driver-agnostic CLI/GUI/MCP. Use Codex MCP for simple second opinions inside Claude Code; use ATO when you need independent multi-reviewer passes, receipts, replay, or the same workflow outside Claude Code. Full write-up: /posts/ato-vs-claude-code-codex-mcp.html
Do ATO and Claude Code + Codex MCP stack?
Yes. Author in Claude Code, run hard multi-model review or war rooms in ATO (which shells out to claude CLI and codex CLI as peers), and optionally let Claude Code call ATO's MCP for ops. They are complementary layers, not mutually exclusive.