Miru: Code Search Built for AI Agents改善
The Challenge
AI coding agents don’t spend most of their tokens writing code — they spend it looking for it.
Grep and glob are cheap per call but expensive in aggregate — an agent using them reads far more of the repo than it needs, and that waste compounds with every retry. Miru targets the ~85% of the token budget that conventional search leaves untouched.
Measured across real tasks, the token bill breaks down like this:
How it Works
Hybrid retrieval, chunked on meaningful boundaries
Miru combines semantic search — via ds1-code, the code-specialised model in the DS1 embedding family — with classic lexical search (BM25), then fuses and reranks the results. Embeddings capture intent; BM25 nails exact tokens like a symbol name. And because ds1-code runs on CPUs we have no limits on usage.
WHAT IT DELIVERS — MEASURED
Miru vs. grep, end to end, across 7 coding scenarios
Grep looks cheap per call, but the agent compensates by reading more, retrying more, and writing more. Those costs only show up when you measure the whole task.
In Takara’s own benchmark, an agent equipped with Miru was run end-to-end against the same agent using grep, across seven coding scenarios. Miru came out ahead on every dimension that matters.
WHERE MIRU FITS
Built for agents, teams, and regulated codebases alike
DIAGRAM GOES HERE
DEPLOYMENT, PRIVACY & SOVEREIGNTY
CPU-only. Cloud, on-premise, or air-gapped.
Because ds1-code is CPU-only, Miru's retrieval needs no scarce GPUs and can run anywhere standard compute is available. The open-source CLI embeds through Takara's hosted API by default; for sensitive code, the same architecture supports local/VPC and air-gapped deployment — code never leaves your perimeter.
LOZENGES GO HERE
GETTING STARTED
Three commands to wire Miru into your agent
bun add -g @takara-ai/miru-code # install
miru setup # store your Takara API key
miru install # wire into your IDE / agent
miru search "auth middleware" ./src
CTA
Book a demo view on GitHub