Coding tools

hanzo --<agent> starts the coding agent you already use, in the repository you are standing in, with the Hanzo toolset attached. The agent keeps its own login and plan. The toolset lets it hand work to Hanzo's models: Enso for text, Kai for typed decisions, plus code and web search. That work is billed to your Hanzo plan, not the agent's.

curl -fsSL hanzo.sh | sh      # hanzo, hanzo-mcp, dev and zt: native binaries, checksummed, into ~/.local/bin
hanzo                         # first run: sign in or sign up, pick your agent, start it
hanzo --claude                # or any agent below
hanzo --claude "fix the failing test in api/users"   # one quoted task runs headless

Sign up and pay

1. Sign up. hanzo on a new machine opens the sign-in picker. Choose Hanzo and the browser opens hanzo.id, where a new account is created. hanzo login does the same at any time. You can also sign in with the agent's own account (Claude, ChatGPT, Google, Cursor); the agent then runs on that account.
2. Pick a plan. hanzo billing plans lists them with current prices. The Free plan includes Hanzo's free models, enso-free and zen-free. Pro, Max 5x, Max 20x and Team include premium and Hanzo model usage each month. Plans are also at hanzo.ai/pricing.
3. Pay. Subscribe at hanzo.ai/pricing or with hanzo billing subscribe card. Add prepaid credit with hanzo billing topup. hanzo billing tier shows your plan and what is left of it, and hanzo billing balance shows prepaid credit.

Every agent

"Added once" means the agent takes no MCP server for a single run, so you add hanzo-mcp to its own configuration one time: gemini mcp add -s user hanzo ~/.local/bin/hanzo-mcp, grok mcp add --scope user hanzo ~/.local/bin/hanzo-mcp, agy mcp add hanzo ~/.local/bin/hanzo-mcp, or Cursor's ~/.cursor/mcp.json (Desktop apps and editors). Each agent is also hanzo dev --<agent> and hanzo dev --backend <agent>, and hanzo code is hanzo dev. hanzo upgrade <agent> installs an npm-shipped agent under ~/.hanzo, where hanzo finds it without changing your PATH. Make one agent the default for a bare hanzo with agent in ~/.hanzo/settings.json.

Hand work to Enso and Kai

The Hanzo toolset is the native hanzo-mcp server, attached to the agent for that run only. Your own agent configuration is never edited. An agent on its own plan can use it to hand work off:

• llm: a request through Enso, Hanzo's router. Enso picks the model your plan pays for and says which one answered. Long reads, summaries and drafts can go here instead of the agent's own model.
• kai_decide: typed questions (yes/no, choice, score) answered by Kai, Hanzo's decision model. Each answer comes with calibrated probabilities and its server latency. Classification, routing and checks over many items fit here.
• code_search, code_context, web_search, web_read, research: search and reading on api.hanzo.ai.
• fs, exec, git, code, lsp, browser: local tools that run on your machine.

Tell the agent which jobs to hand off, or name the tool in your task: "use kai_decide to sort these 200 tickets by queue".

Run the whole session on Hanzo

--cloud sends the agent's own model calls through api.hanzo.ai as well. The model is enso-auto unless you name one with --model. The session then uses your Hanzo plan and none of the agent's.

hanzo --claude --cloud
hanzo --opencode --cloud --model enso-ultra "refactor the store"

What a session used

When a session ends, hanzo prints one block:

hanzo dev · session summary
  session     claude · 16s
  model       claude on its own account; Hanzo bills only what its tools handed off
  hanzo tools 2 calls · 9.5 s
              1 call handed to enso-auto · in 113 · out 373 · 8.4 s each
              1 call handed to kai · in 7 · out 0 · 1.1 s each

• hanzo tools lists every toolset call in the session. Each call that reached a model is named by that model, with the tokens and time its answer reported. A call that reached no model ran on your machine.
• model says whose account the agent's own model calls used. On --cloud the block adds what those calls used: the tokens the agent reported, priced at GET https://api.hanzo.ai/v1/models and labelled as an estimate, or the cloud's meter for the session window.
• hanzo console (press 6) and hanzo usage summary --range 7d show the same usage across sessions, as does platform.hanzo.ai.

Figures are labelled with their source and are never added across sources. Nothing in the block claims a speedup.

Desktop apps and editors

Claude Desktop, Cursor, VS Code, Windsurf, Zed, JetBrains and the Codex and Claude Code apps take the same toolset as an MCP server, and can run their own models on your Hanzo plan: Desktop apps and editors.

How Kai and the toolset measure

Every number Hanzo publishes about Kai and the toolset is a research run, with each item stored beside it.

• Hanzo Dev against Claude Code, 44 held repair tasks (board): Claude Code solved 44 at a 9.6 s median with 3.8 model calls a task; Hanzo Dev solved 40 at 2.2 s with 1.0. On the 40 both solved, Hanzo Dev was 3.8× faster, and 8.2× on the 24 that were search.
• The toolset's local tools (command, select, triage): the command guard answers in 0.06 ms, tool selection in 0.16 ms (top-1 .73), log triage in 0.37 ms (.70 right), each where the model call it replaced took 0.5 to 1.6 s.
• Kai (board): 19.8 ms p50 for one decision.

No study yet measures an agent such as Claude Code with the toolset against the same agent without it. hanzo.ai/benchmarks draws them live: Kai against other decision models, and the agent toolset board, where the injection screen, the command guard, tool selection and log triage each sit beside the handler they replaced, every table linking its runs.

Keep everything up to date

hanzo upgrade                 # hanzo, hanzo-mcp, dev and zt, and every coding agent installed
hanzo upgrade mcp qwen claude # just these

hanzo upgrade runs hanzo.sh into the directory hanzo lives in, so the build that answers hanzo is the one that was upgraded. Each agent upgrades its own way: an npm agent is installed again under ~/.hanzo, one with an updater of its own runs it (claude update, codex update, cursor-agent update, agy update, goose update, grok update, hermes update, interpreter update), and Aider and OpenHands are upgraded with uv tool upgrade. With nothing named, every agent installed that way is upgraded. Every Hanzo tool is a native binary for Linux and macOS on amd64 and arm64, and for Windows on amd64. On Windows, npm install -g @hanzo/cli installs hanzo, and hanzo upgrade fetches hanzo-mcp, dev and zt beside it, each checked against its published sha256.

Options that apply to every agent

hanzo --claude --ask       # the agent asks before each action (default: it acts without asking)
hanzo --claude --no-mcp    # no Hanzo toolset
hanzo --claude --no-link   # the session is not shown in the cloud
hanzo --claude -- --verbose   # everything after -- goes to the agent unchanged

A repository's own .mcp.json and agent settings do not load, because a repository is untrusted and anything it declares would run beside your credentials. --project-mcp opts in for a repository you trust.