HiPMMCode

English

简体中文

English

简体中文

Appearance

Overview ​

hipmmcode is a provider-agnostic AI coding agent that lives in your terminal. It combines a fullscreen terminal UI, a 40+ tool runtime, and a multi-agent orchestration layer, and it can talk to a large catalog of model channels — Anthropic, OpenAI, Gemini, DeepSeek, Kimi, Qwen, xAI Grok (API key or device-code OAuth), Codex subscription, and any custom proxy — through four native wire protocols plus the Claude Code process bridge.

Current release: CLI v1.3.1, desktop 1.2.0. The two version lines stay independent. hipmmcode update and GitHub Latest update the CLI only. The desktop installer is still tag desktop-v1.2.0 and bundles CLI v1.2.0. This desktop release has no Windows installer. macOS is Developer ID signed and not notarized. Linux is unsigned. On a Grok channel, an unset reasoning level defaults to high; the menu is Extended (xhigh), Heavy (high), Balanced (medium), and Faster (low). /login binds a platform plan. The desktop app sends files to the phone and shows meeting notes. DeepSeek images still go through the Files API, and deepseek-flash reads them on the selected model.

bash
hipmmcode                       # fullscreen interactive REPL
hipmmcode "fix the failing test"  # one-shot headless run
hipmmcode serve                 # HTTP/WS server + async task queue
hipmmcode acp --listen :9910    # drive it from an editor over the network

What makes it different ​

Provider-agnostic to the coreTop + Extended channel catalog; each speaks its native streaming protocol (OpenAI-compatible, Anthropic, Gemini, Responses). OAuth for Codex / xAI / Qwen. A custom endpoint is 3 lines of JSON. One agent team can mix models from different vendors.
Fullscreen TUI by defaultA managed viewport with mouse support: hover and click to expand tool output, drag-select with auto-copy, a pinned bottom status bar that survives streaming, an in-turn todo tree, and a transcript view on Ctrl+O.
Multi-agent orchestrationThe Agent tool spawns teammates — foreground or background, on any provider — that coordinate through shared task lists and file mailboxes. Bundled profiles (team-designer, judge, optimizer, Explore, Plan, …) work with zero setup.
Governed autonomyA permission rule engine (allow / deny / ask), everyday permission modes including Auto (new-install default), conditional bypass, lifecycle hooks, an optional OS-level bash sandbox, and hard blocks on destructive commands.
Embeddable everywhereHeadless exec with text / json / stream-json output, a direct-connect HTTP/WS server with a fire-and-forget task queue, the Agent Client Protocol (ACP) for editors, and a stream-json SDK control protocol on stdin/stdout.
Self-documenting binaryhipmmcode integration prints the full machine-integration handbook; hipmmcode doctor diagnoses your environment; hipmmcode update self-updates.

What's new in v1.3.1 ​

What changed in the CLI from v1.2.0 to v1.3.1, grouped by what it is for:

  • Remote control from your phone. /rc (through the platform) and /rcl (HM Bot on your phone over SSH to this computer) let the phone carry on a session from the computer: live replies, permission approvals, images and PDFs, file browsing, a terminal. What was said before turning it on goes along too. /rcs manages the remote state of every session on the computer, and turns it on for a terminal that has it off. See Remote control.
  • Panels inside the fullscreen terminal. Resume, rewind, import and export, plus code review, diff, model, effort, theme, hooks, memory and permissions all open from slash commands without leaving fullscreen. A restart returns to the same session.
  • How a turn looks. A colour dot-matrix HM Bot welcome screen; a spinner line that says what it is doing, with time, tokens and thinking state; Markdown that matches Claude Code, with unordered lists drawn as • ◦ ▪ by level; nothing jumps while a reply is written; CJK text no longer leaves one character alone on a line. See Interactive mode.
  • ultrathink. Write ultrathink in a message and that turn runs at xhigh.
  • Windows. Ctrl+V pastes images in the terminal; selecting text copies it on release; PowerShell scripts with Chinese no longer come out garbled.
  • Works without a VPN. A local proxy left in your shell (HTTPS_PROXY=http://127.0.0.1:7890) is skipped while it is down; a third-party Grok gateway goes in XAI_BASE_URL. See Model channels.
  • Permissions and commands. permissions.blockWritesOutsideWorkingDirectories stops writes outside the working directory; a command that reaches its timeout moves to the background; MCP gains progress, elicitation, list refresh, resource templates and argument completion.

The 30-second tour ​

  1. Install — one curl line, Homebrew, npm, or a PowerShell one-liner on Windows. See Installation.
  2. Connect a model — hipmmcode model walks you through picking a channel, pasting an API key, and choosing a default model. See Quickstart.
  3. Chat — hipmmcode opens the fullscreen REPL. Type naturally; the agent reads files, runs commands, searches the web, and edits code with your approval.
  4. Scale up — ask it to spawn named Agents in parallel; they automatically share the session's single implicit team. Then wire in MCP servers, install skills, define hooks, or drive it headlessly from scripts and servers.

Feature map ​

Platforms ​

macOS (Apple Silicon & Intel), Linux (x64 & arm64), and Windows (x64, native) — distributed as prebuilt binaries. The binary is free to use; see the LICENSE shipped with each release.