Why Anvaya · THE MECHANISMS

Every Other Agent IsHeavy And Alone.Anvaya Is Neither.

Claude Code idles at 432 MB, Codex at 97 MB, Aider at 249 MB. At those weights a laptop cannot host a fleet and a cloud account is the only way to scale. Anvaya measured 25.6 MB median RSS, so hundreds of agents fit on one machine, and the whole thing runs locally.

The Mechanisms

Four Things That Are Hard To Bolt On Later.

01It is measured light25.6 MB median harness RSS across 238 runs in Phase 1, 21 MB floor, with 228 of those runs externally verified. Claude Code idles at 432 MB. That gap is not a slogan; it is the reason the rest of this page is possible.

02It is dense by design24 agents run concurrently by default with a 256-agent hard cap. A 32-agent soak cost 103 MB total. 25 sessions ran in parallel on a Core 2 Duo with 4 GB of RAM, no crashes, no faults. Hundreds on a modern laptop, all local.

03It learns like an engineer, not a cacheOther tools give an agent a memory layer: a place to store facts. Anvaya simulates the lifecycle a junior engineer actually goes through, capturing debugging sessions, dos and don'ts, domain knowledge and project history, so the agent compounds toward seniority instead of re-reading the same files.

04It runs on your machineLocal-first by design. No account, no API key, no per-token bill on the Ollama path, and no requirement that your codebase leave the building. Anthropic and OpenAI-compatible endpoints are there when you opt in, not because anything breaks without them.

Light Enough To Multiply.

A harness that costs 432 MB is a harness you run one of. A harness that costs 25.6 MB is one you run a hundred of, on the machine you already own. Weight is not a vanity metric here; it is what makes local agent density possible.

25.6 MBMedian harness RSS, Phase 1, 238 runs, 228 verified

17xLighter than Claude Code, measured on the same run

256Concurrent agents, hard cap, 24 by default

~1.5 GBApproximate memory for a hundred agents, modern machine

The Junior Engineer Who Never Forgets.

A new engineer does not learn a project from the files alone. They learn it from the debugging session that took all afternoon, the decision nobody wrote down, the domain knowledge the repo never states. Anvaya is built to simulate that lifecycle: not just the codebase, but the experience of working in it, accumulating toward seniority instead of resetting each session.

× Without Anvaya

Session 1: “We decided to use JWT for auth.”
Session 2: “Wait, which auth approach?” [re-reads 30K tokens]
Session 3: “Let me check what changed in auth.” [re-reads 28K tokens]
Session 47: Still re-explaining the same architecture.

You are the project manager AND the institutional memory.

&check; With Anvaya

Session 1: Decision captured as a typed node.
Session 2: Mind injects it automatically, 2K tokens, not 30K.
Session 3: Bug fix recorded; linked to the decision.
Session 47: Mind knows every decision, every failure, every pattern.

The model has institutional knowledge. You focus on building.

What Anvaya Isn’t.

If you’re evaluating this seriously, these will matter more than the feature list.

Pre-1.0 and unreleased. You build the source from scratch; there is no installer, no Homebrew tap, no crates.io publish yet.

Experience has to accumulate. On a fresh project the graph is empty and Anvaya behaves like any other tool; the compounding only starts once sessions build up.

Mind is a Unix-socket daemon. Linux and macOS are the supported surface today.

We do not yet claim the title of world's lightest harness. Phase 1 measured anv in --no-mind mode at the lowest memory on that corpus, models and machine, but the claim needs a shipped default mode, a default-configuration comparison, more harnesses and a non-macOS rerun. What we state is exactly what was measured.

The token-compression headline is a design estimate. The measured end-to-end lever today is -11.0% median uncached tokens (micro-fold, n=5, verification 19/19); your ratio depends on how much of your work repeats.

Run Agents That Fit On Your Laptop.

25.6 MB median RSS. 25 agents ran in parallel on a Core 2 Duo with 4 GB RAM. Hundreds on your machine. Zero cloud required on the Ollama path.

Get Started →Read the FAQ

Requires Rust/cargo to build from source. Linux and macOS today, Windows not yet supported. Pre-1.0, public beta. Pricing TBD.