mimo-tui

A native MiMo terminal agent with real tool use, MCP support, sessions, streaming, and no Claude-Code-style ToS gray area

5-min Quickstart GitHub

v0.3.0-alpha.2Rust · 6 platformsMIT LicensedFree Forever

Terminal-native product

Not a dressed-up docs page. A real product surface for MiMo.

mimo-tui is built around MiMo’s actual model behavior: streaming, thinking, tool loops, session management, cache visibility, and multi-region fallback. The goal is simple: make MiMo feel production-ready in the terminal.

Original architectureReal-key verifiedMCP dual protocolParallel sub-agents

What shipping looks like

102docs pages

8built-in tools

3built-in skills

32max tool iterations

From install flow to sandboxing, this is structured like a maintained CLI product, not a thin experiment.

Get working in three steps

Install it

npm, brew, cargo, docker, and curl are all supported. No repo archaeology required.

Set up your key

The wizard handles MiMo Token Plan keys, region selection, and fallback behavior cleanly.

Start shipping

Once inside the TUI, you can edit files, run shell commands, attach MCP servers, and split work across agents.

npm install -g mimo-tui
mimo init
mimo

brew install duolaAmengweb3/mimo-tui/mimo-tui
mimo init
mimo

cargo install --git https://github.com/duolaAmengweb3/mimo-tui mimo-tui
mimo init
mimo

curl -fsSL https://mimo-tui.pages.dev/install.sh | sh
mimo init
mimo

docker run -it -v "$PWD":/workspace \
  -e MIMO_API_KEY=tp-xxx \
  ghcr.io/duolaamengweb3/mimo-tui:latest

Terminal UX should feel like a product, not a log stream

Interaction design

Expose the work, not just the answer

thinking, tool calls, type checks, cache hits, and session state are all visible in real time. That gives users observability into what the agent is actually doing instead of hiding everything behind one opaque reply.

Live streaming and thinking rendering
Tool execution and failure reasons stay visible
Prompt-cache savings are surfaced directly
Slash commands, sessions, and modes stay inside the terminal workflow

∞ mimo-tui v0.3.0-alpha.2 · mimo-v2.5-pro · SGP · agent

▎ user > Refactor Header.tsx to a function component, preserve props, then run typecheck

  thinking find file → read original → inspect prop types → rewrite → verify
  ✓ read_file {"path":"src/components/Header.tsx"}
  ✓ edit_file {"path":"src/components/Header.tsx", ...}
  ✓ shell {"command":"npm run typecheck"}
    no errors found

Header.tsx is now a function component. Type-check passed.

  usage in:1842 out:312 · cached 1568

Core capabilities beyond “it can call an API”

Complete agent loop

Streaming SSE, tool execution, and up to 32 iterations, enough for non-trivial coding tasks.

8 built-in tools

read / write / edit / shell / glob / grep / web_fetch / todo out of the box.

Parallel sub-agents

The task tool can dispatch up to 10 concurrent sub-agents for decomposition-heavy work.

MCP dual-protocol

Both stdio and HTTP are supported, so existing MCP servers plug in with minimal friction.

Skills-compatible

Existing .claude/skills/ content can be migrated directly with very little ceremony.

Three-platform sandbox

macOS Seatbelt, Linux Landlock, and Windows Job Objects are built in from the start.

Plan / Agent / Auto modes

Switch between read-only, approval-based, and YOLO execution depending on risk tolerance.

Smart regional routing

CN / SGP / AMS latency checks plus automatic fallback when a key does not work in the chosen region.

The important difference is product judgment, not just visuals

	Claude Code wrapper	XIAOMI-MiMo-code	mimo-tui
Architecture	Anthropic shell	Reverse-engineered	Original implementation
Model identity	“I’m Claude”	Prompt patching	Native MiMo identity
Config directory	Collision-prone	Isolated	Dedicated `~/.mimo/`
Thinking / cache visibility	Partial	Incomplete	Fully exposed
MCP support	Yes	Partial	stdio + HTTP
Compliance	ToS gray area	High risk	Original + MIT
Maintainability	Upstream constrained	Fragile	Can evolve independently

Proven, not hypothetical

supported platforms

macOS arm/x64 · Linux x64/arm64 gnu+musl · Windows x64

distribution channels

npm · brew · cargo · docker · curl

354K+

cached tokens observed

Measured across 75 real API calls, not estimated marketing math

102

bilingual docs pages

Install, MCP, skills, sandbox, usage, troubleshooting, and more

Real coding flow verified

mimo -p "write greet.py and run it" exercises file writing, shell execution, and Python-style skill behavior.

MCP is live

The filesystem server’s 13 tools were auto-registered and called successfully.

Fallback actually works

The wizard has already validated CN 401 failure followed by a successful switch to SGP.

Sandboxing is enforced

macOS Seatbelt blocked out-of-sandbox writes, proving the safety boundary is operational.

Who this is for

For builders

People who want MiMo inside their real dev workflow

If you do not want to babysit prompts, wrappers, and environment-variable hacks, this is the sharper path.

For teams

People who care about maintainability and clean separation

Independent config, original code, and explicit protocol support make future extension much less fragile.

Entry points

GitHub

Source and issues

duolaAmengweb3/mimo-tui

Bilingual · 102 pages

Read docs

Free, and intended to stay that way

mimo-tui is MIT licensed. No Pro wall, no enterprise upsell layer, no ads, and no managed lock-in.

Users bring their own MiMo Token Plan API keys, so billing stays between the user and the model provider.
The tool layer stays light and original, with the focus on reliability, protocol compatibility, and workflow quality.
Future work should keep improving execution quality instead of adding ornamental pricing tiers.

“MiMo” is a registered trademark of Xiaomi. mimo-tui is an independent community project and is not officially affiliated with Xiaomi.