Miru code agent

Miru (見る)

Hybrid code search for AI coding agents. Find code by meaning, not grep.

Your AI agent finds the code it needs with up to 50% fewer tokens.

Miru returns the best chunks (path, lines, snippet) for questions like "where is auth middleware configured?" — plugged directly into Claude Code, Cursor, Copilot, Codex, and 9+ other agents via MCP.

---

How it fits into your agent's workflow

Miru replaces the grep/glob-style search agents fall back on today. Install once, and every connected agent gets search and find_related MCP tools automatically.

Install via a plugin marketplace

Prefer this when your IDE has a plugin marketplace — no bun add -g step, and updates go through the IDE's own plugin flow. Run miru setup first if you have not authenticated yet (see Set up credentials).

Claude Code:

/plugin marketplace add takara-ai/miru-code
/plugin install miru

Restart Claude Code (or reload plugins) when prompted. Update: /plugin marketplace update miru then reinstall. Remove: /plugin uninstall miru.

Cursor:

1. Dashboard → Plugins → Team Marketplaces → Add Marketplace → Import from Repo → takara-ai/miru-code
2. Customize (sidebar) → find miru → Install → choose project or user scope

Any other IDE, or if marketplace install isn't available: use the CLI install below.

What you get from a plugin install vs the CLI

Plugin installs don't carry full CLI parity — what each one gives you depends on the IDE's own plugin capabilities, not just Miru's packaging:

	Claude Code	Codex	Cursor
MCP tools (search, locate, expand, find_related)	✅	✅	❌ (plugin ships skills + rules only — no MCP entry yet)
miru / Caveman / STE skills	✅	✅	✅
Dedicated sub-agent (miru:miru-code)	✅	—	—
Benchmark mode toggle (/plugin configure)	✅	—	—
Credentials	own plugin-scoped dir	own plugin-scoped dir	n/a

— means the IDE's plugin schema has no equivalent mechanism to port these to (not a packaging gap we can close): Codex's and Cursor's plugin manifests have no agents or userConfig fields, so the dedicated sub-agent and the benchmark toggle are Claude-Code-only.

Credentials are plugin-scoped, not shared with your CLI install. Claude Code and Codex plugins each store auth in their own IDE-managed data directory (survives plugin updates, removed on uninstall). This means miru setup done via the CLI does not carry over to a plugin install, or vice versa — each authenticates independently on first use (interactive device-code login bootstraps automatically). If you use both the CLI and a plugin, you'll sign in twice.

Install (CLI)

bun add -g @takara-ai/miru-code

Set up credentials

miru setup

Interactive miru setup defaults to device-code login and saves the resulting credentials locally. Manual bearer-token entry is still available with --key. If credentials are missing, the interactive MCP/plugin path can bootstrap the same device flow automatically on first use.

miru setup --device           # explicit device-code login
miru setup --key YOUR_TOKEN   # store a bearer token directly
miru setup --clear            # remove stored credentials

Miru stores versioned credentials in credentials.json and automatically loads or refreshes them for MCP and CLI use. TAKARA_API_KEY still overrides stored credentials when set explicitly.

Add to your IDE

miru install

Interactive TUI — ↑↓ move, space toggle, a all, enter confirm. Pick agents and integrations:

Integration	What it does
MCP server	search, locate, expand, and find_related tools in the agent
Instructions	Search policy in CLAUDE.md / AGENTS.md / GEMINI.md
Sub-agent	Dedicated miru-code agent file
Cursor rules	Always-on .cursor/rules/miru-code.mdc (Cursor only)
Old Miru hooks	Removes stale search hooks from older releases
Caveman (experimental)	On-demand chat compression skill (/caveman)
STE writing (experimental)	On-demand clear technical English for docs (/ste)

Restart the IDE when done.

miru uninstall   # remove miru config

Supported: Cursor · Claude Code · Gemini CLI · Kiro · OpenCode · GitHub Copilot · Codex · VS Code · Visual Studio (Windows) · Windsurf / Devin Desktop

IDE	MCP	Instructions / rules	Caveman (experimental)	STE (experimental)
Cursor	~/.cursor/mcp.json	~/.cursor/rules/miru-code.mdc	~/.agents/skills/caveman/SKILL.md	~/.agents/skills/ste/SKILL.md
Claude Code	~/.claude.json	~/.claude/CLAUDE.md	~/.claude/skills/caveman/SKILL.md	~/.claude/skills/ste/SKILL.md
Gemini CLI	~/.gemini/settings.json	~/.gemini/GEMINI.md	~/.agents/skills/caveman/SKILL.md	~/.agents/skills/ste/SKILL.md
Kiro	~/.kiro/settings/mcp.json	~/.kiro/steering/miru.md	~/.kiro/skills/caveman/SKILL.md	~/.kiro/skills/ste/SKILL.md
OpenCode	$XDG_CONFIG_HOME/opencode/opencode.json(c) (else ~/.config/opencode/…)	…/AGENTS.md	~/.agents/skills/caveman/SKILL.md	~/.agents/skills/ste/SKILL.md
GitHub Copilot	~/.copilot/mcp-config.json	—	~/.agents/skills/caveman/SKILL.md	~/.agents/skills/ste/SKILL.md
Codex	~/.codex/config.toml	~/.codex/AGENTS.md	~/.agents/skills/caveman/SKILL.md	~/.agents/skills/ste/SKILL.md
VS Code	…/Code/User/mcp.json	—	~/.agents/skills/caveman/SKILL.md	~/.agents/skills/ste/SKILL.md
Visual Studio	%USERPROFILE%\.mcp.json	—	~/.agents/skills/caveman/SKILL.md	~/.agents/skills/ste/SKILL.md
Windsurf	—	—	~/.agents/skills/caveman/SKILL.md	~/.agents/skills/ste/SKILL.md

Plugin packaging

• Codex: .codex-plugin/plugin.json, .agents/plugins/marketplace.json, shares the root mcp.json
• Claude Code: .claude-plugin/plugin.json, .claude-plugin/marketplace.json, its own .claude-plugin/mcp.json (needed for the userConfig benchmark toggle — see What you get from a plugin install vs the CLI)
• Cursor: plugin.json and .cursor/rules/miru-code-search.mdc

Current limitation:

• these plugin manifests launch the published Miru runtime through bunx @takara-ai/miru-code@latest mcp
• that means local source edits do not affect plugin behavior until a package version is published
• and a fully self-contained “no Bun required” plugin install is still future work

On-demand skills

Sub-agent files are also written where supported (see miru install plan). Windsurf hooks only (experimental) — no MCP entry yet. Caveman is an on-demand Agent Skill (default off): invoke with /caveman or “talk like caveman”; stop with “normal mode”. Invocation UI varies by IDE (/caveman, $caveman, @caveman, etc.). Most IDEs share ~/.agents/skills/caveman/SKILL.md (including Copilot / VS Code / Visual Studio); Claude Code and Kiro keep vendor-native skill dirs. Ownership is tracked on the shared path so uninstalling one IDE keeps the skill while another still owns it; selecting all owners removes it once. STE is an on-demand Agent Skill (default off): invoke with /ste or “de-slop this”; keep articles and complete sentences. Most IDEs share ~/.agents/skills/ste/ the same way (ownership via miru-owners.json); Claude Code and Kiro keep vendor-native STE dirs.

STE writing (experimental)

STE helps write clear technical English for docs, runbooks, errors, and release notes (pragmatic ASD-STE100-inspired rules). Invoke with /ste or “de-slop this”. Keep articles and complete sentences — not telegraph-style omission.

Not ASD-certified. Full dictionary compliance needs the official standard at asd-ste100.org. Miru does not ship the copyrighted ASD dictionary.

Not for marketing or brand copy. Off by default; enable STE in the installer for any supported IDE. Restart the IDE (or reload skills) after install. For Codex, install also sets [features] skills = true in ~/.codex/config.toml (same as Caveman).

Caveman mode (experimental)

Caveman compresses live chat replies (less filler, max meaning). Intensities: /caveman lite|full|ultra (default full). Persisted artifacts (commits, PRs, customer docs) stay normal prose unless you ask otherwise.

Security / destructive warnings use clear normal prose (auto-clarity) — brevity never hides risk. Session token savings vary; the skill itself costs input tokens. No guaranteed %.

Off by default at install time. Enable Caveman in the installer for any supported IDE. Restart the IDE (or reload skills) after install. For Codex, the installer also sets [features] skills = true in ~/.codex/config.toml (required for Codex to load skills).

Team sub-agent in a repo (optional):

miru init --agent claude --force

Try it

Meaning-based questions → search. Exact strings (env vars, symbols, error codes) → locate.

miru search "auth middleware" ./src
miru locate REDIS_HOST ./src
miru expand src/auth.ts 42 ./src
miru find-related src/auth.ts 42 ./src

Terminal output is human-readable; use --json for scripts. One-off without installing:

bunx @takara-ai/miru-code@latest search "auth middleware" ./src

---

MCP tools

When wired via miru install, the MCP server exposes search, locate, expand, and find_related. read_benchmark appears only in benchmark mode. repo is optional: it defaults to the server's startup directory. Set it for another repo, or if your MCP client starts elsewhere. The index is built on the first call and cached for the session.

Tool	When to use
search	Default for code exploration — hybrid semantic + keyword search. One call per question.
locate	Exact substrings (env vars, symbols, error codes) — prefer over Grep.
expand	More context in the same file when a hit has truncated: true.
find_related	Similar code in other files from a file_path + anchor_line.
read_benchmark	Cumulative token-savings rollup (benchmark mode only).

Workflow

1. search with query — returns compact snippets (~±15 lines) and relevance scores.
2. If a hit has truncated: true, call expand with file_path and anchor_line — not another search or a full-file read.
3. To trace similar patterns elsewhere, call find_related with the same file_path and anchor_line.
4. Use your editor's Read on absolute_path only when editing or when expand still lacks context.

Keep repo on follow-up calls when set.

Prefer these tools over Grep, Glob, or SemanticSearch when Miru MCP is connected — hooks and instructions enforce that when enabled.

Local repo hits include absolute_path for one-click navigation. Parameter reference is below under MCP parameters.

How it works

Hybrid search: Takara embeddings + BM25 + fusion + reranking. Index code, docs, config, or all with --content.

MCP watches local files and updates the index incrementally. Package upgrades invalidate stale caches via the version epoch.

OS	Index cache
macOS	~/Library/Caches/miru
Linux	~/.cache/miru
Windows	%LOCALAPPDATA%\miru\Cache

Chunking & languages

Miru chunks source in tiers: AST (tree-sitter, default) → structural heuristics → line splits.

AST chunking — 26 languages (syntax-aware boundaries via vendored web-tree-sitter grammars):

Language	Typical extensions
astro	.astro
bash	.sh, .bash, .zsh
c	.c
cpp	.cpp, .h, .hpp, etc.
csharp	.cs
css	.css
dart	.dart
elixir	.ex, .exs
embeddedtemplate	.erb, .ejs
go	.go
haskell	.hs
html	.html, .htm
java	.java
javascript	.js, .jsx, .mjs, .cjs
json	.json
ocaml	.ml, etc.
php	.php
python	.py, .pyi
ruby	.rb
rust	.rs
scala	.scala
solidity	.sol
sql	.sql
svelte	.svelte
typescript	.ts, .tsx, .mts, .cts
vue	.vue

Structural fallback (brace/indent heuristics when AST is unavailable): python, go, typescript, javascript, cpp, c.

Line fallback: everything else that gets indexed (kotlin, swift, etc.) — still searchable, coarser chunks.

Set MIRU_AST_CHUNKING=0 to disable AST and use structural → lines only.

CLI reference

Run miru in a terminal or miru -h for the command list, or miru <command> -h for details.

Command	Purpose
miru setup	Authenticate and store credentials
miru install	Configure IDE (global)
miru uninstall	Remove IDE config
miru search <query> [path]	Search (-k N, --content, --json)
miru locate <literal> [path]	Exact substring in the index
miru expand <file> <line> [path]	Adjacent chunks in the same file
miru find-related <file> <line> [path]	Related chunks
miru benchmark on/off/status/clear	Toggle MCP benchmark mode / clear report
miru init --agent <id>	Project-local sub-agent
miru clear [path]	Drop index cache (use after big CLI-only refactors)
miru mcp	Start MCP server (--benchmark for comparisons)

CLI uses hyphens (find-related); MCP tool names use underscores (find_related).

Library

bun add @takara-ai/miru-code

import { MiruIndex } from "@takara-ai/miru-code";

const index = await MiruIndex.fromPath("./src");
const results = await index.search({ query: "BM25 tokenize" });

Environment

Variable	Notes
TAKARA_API_KEY	Required
MIRU_OPENAI_BASE_URL	Default https://infer.takara.ai/v1
MIRU_OPENAI_EMBEDDING_MODEL	Default ds1-miru-int8
MIRU_WORKSPACE_ROOT	Optional: restrict MCP local repo paths to this directory
MIRU_MAX_INDEX_FILES	Cap files indexed per operation
MIRU_ALLOW_HTTP_GIT	Set 1 to allow plain http:// git clones
MIRU_MCP_WATCH	Set 0 to disable MCP file watch
MIRU_AST_CHUNKING	Set 0 to disable tree-sitter AST chunking
MIRU_BENCHMARK_HISTORY_PATH	Override; see Benchmark mode
MIRU_CACHE_HOME	Override the platform cache root (see How it works)
MIRU_QUIET	Set 1 to skip the framed CLI banner (subtitle only on color terminals)
NO_COLOR	Disable CLI colors

See .env.example for more.

Privacy and API usage

Miru sends file contents to the Takara inference API when building an index and when embedding search queries. Chunks from your repo are transmitted over HTTPS to generate embeddings. API usage may incur cost depending on your Takara plan.

If you index proprietary code, make sure that sending snippets to Takara's endpoint fits your security and compliance requirements. MIRU_WORKSPACE_ROOT is an opt-in boundary for MCP local repo paths only, and restricts indexing to a single workspace directory when set.

Enterprise self-hosted embeddings (no Takara egress): see docs/self-hosted-sagemaker.md.

Benchmark mode

Optional measurement of how many tokens Miru saves versus a simple Grep workflow (ripgrep + reading the top matched file). Useful when evaluating Miru; leave it off day-to-day. Only local repo paths are compared — git URL repos skip the comparison and return benchmark_skipped: "local_repo_only".

miru benchmark on       # add --benchmark to installed MCP configs
miru benchmark status
miru benchmark off      # prefer off when finished measuring
miru benchmark clear    # delete the global report

Restart agents after changing mode. miru install keeps --benchmark if it was already enabled.

While on, each search / locate response includes a compact benchmark block (save_pct, miru_tok, grep_tok, saved_tok, rank1). Call read_benchmark for a cumulative rollup (or ask the agent when you want totals).

History is global (not per-repo) under Miru's state directory:

OS	Default path
macOS	~/Library/Application Support/miru/benchmark-history.json
Linux	~/.config/miru/benchmark-history.json ($XDG_CONFIG_HOME/miru/… when set)
Windows	%APPDATA%\miru\benchmark-history.json

Override with MIRU_BENCHMARK_HISTORY_PATH. Append-only JSONL of compact token deltas (no query text); read_benchmark returns cumulative totals. Stored in plaintext — miru benchmark clear or uninstall on shared machines.

MCP parameters

**search**

Param	Required	Notes
query	yes	Natural language or code query
repo	no	Startup directory by default; set for another repo
include	no	Gitignore-style glob patterns; only matching files are searched (same as locate.include)
exclude	no	Gitignore-style glob patterns; matching files are skipped (same as locate.exclude)
dedupe_by_file	no	Keep best hit per file (default true)

**locate**

Param	Required	Notes
literal	yes	Exact substring to find
repo	no	Startup directory by default; set for another repo
include	no	Gitignore-style glob patterns; only matching files are searched
exclude	no	Gitignore-style glob patterns; matching files are skipped
mode	no	count · locations · lines (default). Prefer count/locations when possible
limit	no	Cap returned hits. Omit to return all matches
ignore_case	no	Case-insensitive match (default false)

**expand**

Param	Required	Notes
file_path	yes	From hit file_path or absolute_path (local repos)
anchor_line	yes	From the search hit (anchor_line when truncated, else start_line)
repo	no	Same repo as the search, if set
before / after	no	Extra chunks before/after anchor (default 1 each)

**find_related**

Param	Required	Notes
file_path	yes	From a search hit
anchor_line	yes	From the search hit
repo	no	Same repo as the search, if set

**read_benchmark** (benchmark mode only)

Param	Required	Notes
repo	no	Filter rollup to one local path or git URL. Omit for all saved queries

Manual MCP (skip miru install)

{
  "miru": {
    "command": "miru",
    "args": ["mcp"]
  }
}

For benchmark mode, set "args": ["mcp", "--benchmark"] (or append the flag). Prefer miru benchmark on after a normal install — it updates every agent config.

Run miru setup once so the server can load credentials from credentials.json. If the MCP server starts in an interactive terminal without stored credentials, it will start device login automatically.

Use bunx + @takara-ai/miru-code@latest if miru is not global. The installer uses this command so each new MCP server launch can pick up a published version without a global update. A running server keeps its current version until restarted. Wrapper key varies by IDE (mcpServers, servers, or mcp).

Older headless MCP configs that launch miru without a subcommand, with or without MCP flags, continue to work. Run miru install again to update managed entries to the explicit mcp subcommand, then restart your coding agent. For manual MCP configs, add mcp after the executable or package name.

Developing

git clone https://github.com/takara-ai/miru-code.git && cd miru-code
bun install && cp .env.example .env.local
bun test && bun run typecheck

See CONTRIBUTING.md for pre-commit hooks, commit message conventions, and the PR process.

Local MCP: "command": "bun", "args": ["/path/to/miru-code/src/cli.ts"]

miru -h / setup print a framed wordmark on color terminals (MIRU_QUIET=1 for subtitle only). Crane art lives in src/brand-banner.ts; regenerate with bun run scripts/render-crane-art.ts (ImageMagick). The crane is a registered mark of Takara.ai Ltd.

Credits

Miru uses work by MinishLab. Thank you to its authors.

• semble (MIT). Miru ports parts of semble. These parts include the file walker, the index layout, the BM25 tokenizer, the hybrid search pipeline, and the ranking signals.
• potion-code and Model2Vec (MIT). The WordPiece tokenizer in tokenizer/tokenizer.json comes from potion-code.

See NOTICE for the licence text and the Model2Vec citation.

License

MIT