Miko

简体中文 | English

Miko is a terminal-first AI coding agent specially optimized for the Xiaomi Mimo model, built for developers who want a fast local workflow, a polished TUI, strong model support, and real extensibility without a heavy desktop stack.

Miko is maintained at jianga0801-ui/miko and is based on the OpenCode codebase from sst/opencode. It keeps the parts that make OpenCode powerful - TUI, MCP, providers, tools, sessions, and plugins - and pushes the fork toward a cleaner local-first Miko distribution with bundled skills, cross-platform release binaries, and Windows ergonomics.

Why Miko

• Specially Optimized for Xiaomi Mimo: First-class deep optimizations for the Xiaomi Mimo large model, maximizing cache hits, fully utilizing multimodal inputs (audio, video, image), and guaranteeing premium code generation quality.
• Open the archive and run it: release binaries are self-contained for the core CLI/TUI runtime. End users do not install Bun, Node.js, or node_modules.
• Terminal-native workflow: Miko starts in the project directory, reads the repository, edits files, runs commands, and keeps the loop inside the terminal.
• Model-aware runtime: model limits, modalities, tool support, reasoning, variants, and provider options are normalized before requests are sent.
• Cache-hit optimized agent loop: prompt caching is on by default where the provider supports inline cache hints, reducing repeated tool/system context cost in long tool-use turns.
• Built-in skills and slash commands: project workflows such as review, planning, design polish, document generation, translation, changelog, and cleanup are available without extra setup.
• Extensible by design: MCP prompts, local command templates, plugins, provider hooks, workspace adapters, keymaps, and TUI slash commands are all first-class extension points.

Xiaomi Mimo Model Deep Optimization 🚀

Miko is meticulously tailored and optimized for the Xiaomi Mimo large model, delivering unmatched performance and a state-of-the-art coding experience:

• Cache-Hit Optimization: Advanced prompt-caching strategies designed specifically for Mimo's context window. It dramatically reduces token overhead and latency during long, tool-heavy development loops by keeping system prompts, tool definitions, and repo context cached.
• Full Multimodal Understanding: Fully unleashes Mimo's multimodal capabilities, seamlessly processing and understanding rich inputs including images, audio, and video for visual coding, UI review, and multimodal debugging.
• Elite Reply Quality: Finetuned formatting prompts, system context structures, and response generation rules tailored to Mimo to ensure highly accurate, clean, structured, and production-grade code output.
• Built-in Tavily Search: Deeply integrated Tavily search capabilities (requires a Tavily API Key) to empower the agent to fetch up-to-date documentation, API specs, and web resources for highly precise answers.

Install

Linux and macOS:

curl -fsSL https://raw.githubusercontent.com/jianga0801-ui/miko/main/install | bash
miko --version

Windows PowerShell:

irm https://raw.githubusercontent.com/jianga0801-ui/miko/main/install.ps1 | iex
miko --version

Manual downloads are available from GitHub Releases. Each release ships Linux, macOS, Windows, glibc/musl, x64/arm64, baseline builds, and checksums.txt.

See docs/install.md for version-pinned and manual installs.

What Is Built In

Agents

Miko ships with specialized agents instead of one generic prompt. The specialized built-in agents and workspaces in Miko are inspired by and refer to opencode-workspace:

Agent	Role
miko	Read-only orchestrator that routes work to focused subagents.
coder	Implementation specialist for source edits, builds, tests, and fixes.
explore	Read-only codebase explorer for tracing files and logic.
researcher	External research agent for docs, APIs, package data, and web sources.
reviewer	Code and plan reviewer with severity and confidence thresholds.
scribe	Documentation, changelog, release note, PR, and user-facing prose writer.
.miko/agent/triage	GitHub issue triage agent for ownership routing.
.miko/agent/duplicate-pr	GitHub PR duplicate detector.

Skills

Built-in skills can be loaded by agents and are also exposed as command-style workflows when appropriate:

Skill	Purpose	Source
code-philosophy	Internal logic and data-flow standards.	Miko builtin, this repository.
frontend-philosophy	UI and visual quality standards.	Miko builtin, this repository.
code-review	Four-layer review process: correctness, security, performance, maintainability.	Miko builtin, this repository.
plan-protocol	Implementation-plan format, citations, and progress tracking.	Miko builtin, this repository.
plan-review	Quality checks for implementation plans.	Miko builtin, this repository.
effect	Effect v4 / effect-smol coding guidance for this repo.	Local project skill; API reference source is Effect-TS/effect-smol.

Plugins

Miko's TUI is powered by a set of built-in feature plugins that load automatically at startup with zero configuration. They share the same plugin interface as third-party plugins (see "Extensibility" below), so you can author and install your own the same way.

Plugin	Purpose
which-key	Keybinding hint overlay showing available shortcuts and pending key sequences.
diff-viewer	Built-in diff viewer with file-tree navigation and split-pane comparison.
plugin-manager	Plugin management UI to browse and enable / disable installed plugins.
notifications	Desktop / system notifications when a task finishes or needs attention.
sidebar-context	Sidebar: current context usage and session info.
sidebar-mcp	Sidebar: connected MCP servers and their status.
sidebar-lsp	Sidebar: LSP server status (enabled by default).
sidebar-todo	Sidebar: current task / TODO list.
sidebar-files	Sidebar: files changed during the session.
home-footer, home-tips, sidebar-footer	Footer info and usage tips on the home screen and sidebar.

Slash Commands

Miko features an elegant and highly customizable slash command system. Commands can come from built-ins, project configuration files, MCP server prompts, or custom skills.

Built-in commands:

Command	What it does
/init	Guided AGENTS.md setup for a repository.
/review	Delegates code review to the reviewer agent.

Project commands in .miko/command:

Command	What it does
/ai-deps	Reports safe AI SDK dependency upgrades.
/changelog	Builds UPCOMING_CHANGELOG.md from structured release input.
/commit	Creates and pushes a conventional commit from current diffs.
/goal	Runs a high-thoroughness goal execution protocol.
/issues	Searches GitHub issues for related reports.
/learn	Extracts durable repo learnings into scoped AGENTS.md files.
/plan	Researches and saves a cited implementation plan.
/rmslop	Removes AI-generated code/documentation slop from a branch.
/spellcheck	Checks changed Markdown for spelling and grammar.
/translate	Translates changed English docs and UI copy while preserving code terms.
/yolo	Direct execution mode with verification and no planning detour.

All active MCP prompts and custom skills are automatically registered as slash commands (/), providing a unified and autocompletable TUI shortcut experience.

Model and Provider Runtime

Miko combines two model layers:

• models.dev catalog data for provider/model metadata, refreshed with a local disk cache and an in-process cache.
• Runtime provider adapters that normalize authentication, endpoint selection, request options, streaming events, and model capabilities.

Miko is specifically designed for the Xiaomi Mimo model family as its primary, fully optimized provider.

Miko tracks capabilities instead of treating every model as plain text:

• context, input, and output limits
• text, image, audio, video, and PDF input/output modalities (perfectly supporting Mimo's multi-modal features)
• tool-call support
• reasoning support and reasoning variants (specially optimized for Mimo's reasoning-content continuation)
• temperature support
• provider-specific endpoint type
• cost, cache read/write cost, and status metadata

Provider-specific optimizations include:

• First-class Mimo provider optimizations, integrating token usage tracking, reasoning part separation, and high-performance streaming.

Cache-Hit and Context Optimizations

Miko optimizes repeated agent turns at several layers:

• Prompt cache auto-placement: for providers that respect inline cache markers, Miko places cache breakpoints at the last tool definition, last system part, and latest user message. This targets the stable prefix used repeatedly during a single tool-heavy turn.
• Manual cache hints preserved: explicit cache markers on tools, system parts, or messages are not overwritten.
• Provider-aware behavior: inline markers are applied to Mimo and other compatible paths, while providers with implicit caching or out-of-band cache systems are left alone.
• Usage accounting: usage events keep non-cached input tokens, cache-read tokens, cache-write tokens, reasoning tokens, visible output tokens, and raw provider metadata separately.
• Model catalog cache: models.dev metadata is cached on disk, guarded by a cross-process file lock, and refreshed in the background.
• Local search fallback: release builds no longer depend on downloading ripgrep at runtime; if a system rg is unavailable, Miko falls back to the built-in file search path.

Extensibility

Miko can be extended without patching core code, fully compatible with all original OpenCode extensions:

• MCP: MCP prompts become slash commands; MCP tools become available through the agent runtime.
• Commands: add Markdown files under command/ or .miko/command/.
• Skills: add SKILL.md directories under the configured skills paths.
• Plugins: register tools, provider hooks, model hooks, keymap layers, TUI routes, slash commands, workspace adapters, and context injectors.
• Providers: use bundled AI SDK providers or define custom endpoints and provider-specific options.

Source Development

End users should use the release installers above. Source development requires Bun:

git clone https://github.com/jianga0801-ui/miko.git
cd miko
bun install
bun run dev

Run package checks from package directories. Do not run the root test script; it intentionally exits.

cd packages/miko
bun typecheck

Release

The release workflow builds self-contained archives and uploads checksums:

script/release 0.0.1

The same workflow also runs when a v* tag is pushed.

Fork and Source Attribution

Component	Repository
Current Miko repository	jianga0801-ui/miko
Primary upstream codebase	sst/opencode
Historical package metadata still visible in some files	anomalyco/miko
Model catalog	models.dev
Built-in agent workspace references	kdcokenny/opencode-workspace
Effect reference used by local effect skill	Effect-TS/effect-smol