Pillar · Local Agents

A Coding Agent That Never Sends Your Code Anywhere.

Ollama-first, offline-capable, one small Rust binary. Privacy you can prove, costs that don't scale with usage, and access to the systems behind your firewall.

Direct answer

A local AI coding agent runs its model and its edit-test loop on your own hardware, so source code, prompts, and diffs never leave your machine. Anvaya CLI is built for exactly this: Ollama-first with zero API keys to start (devstral by default), a single ~15MB Rust binary at ~5MB idle, and a local memory daemon under .anvaya/mind/ that remembers decisions, patterns, and bugfixes across sessions. It runs fully offline after one model pull, talks to Anthropic, OpenAI, or a custom gateway only if you configure a key, and stays honest about the tradeoff — local 30B-class models cover everyday work free, while frontier cloud models can be escalated per task for the hardest refactors. Privacy, zero marginal cost, and compounding memory: that is the local-first case.

Why Local

Privacy, Cost, Control.

The three reasons teams go local — and why memory is the fourth.

01Ollama-first, zero keys to startDefault model devstral via Ollama on localhost:11434. Embeddings, summarization, and generation all run locally. Anthropic, OpenAI, and custom gateways exist only if you configure a key — nothing phones home by default.
02Proven privacy postureSource, prompts, diffs, and the memory graph under .anvaya/mind/ never leave hardware you own. The strongest GDPR/data-residency story there is: there is no vendor processing your code at all.
03Genuinely offlinePull the model once and the whole loop — agent, memory daemon, tools — runs without connectivity. Flights, VPN-only networks, air-gapped repos: the use-cases cloud agents simply cannot cover.
04Small enough to auditOne ~15MB Rust binary, ~5MB idle footprint, no Node runtime, no GC pauses. Every tool call and every memory decision is inspectable on your own machine.
05Local-first, not local-onlyEscalate hard tasks to frontier cloud models per session while daily work stays local and free. Route simple work to Ollama, reserve paid inference for the refactors that earn it.

Setup

Local in Minutes.

Install Ollama, pull a model, build Anvaya from source (pre-1.0). Full guide on /start.

011 · Model runtimeInstall Ollama, then pull a tool-calling coder (devstral default). Verify it answers on localhost:11434.
022 · Agent + memoryBuild Anvaya from source (Rust 1.75+), run anv init in your repo — config, sessions, and mind/ scaffold in under 50ms.
033 · Work privateCode, plan, and refactor with the graph compounding locally. Escalate to a cloud provider per task only when the job earns it.

Questions

Asked About Local Agents.

Q

What is a local AI coding agent?

A coding agent whose model and edit-test loop run on your own hardware: inference on your GPU or unified memory, code and prompts never leaving your machine. Anvaya pairs an Ollama model runtime with a Rust agent harness and a local memory daemon.

Q

Can I run a coding agent fully offline?

Yes. Pull the model once (e.g. ollama pull devstral), and Anvaya's agent loop, memory daemon, and tools run with no network. Pre-cache any packages you need; everything else is already on disk.

Q

Which local model should I use for coding?

Whatever fits your hardware with a 32K+ context window and tool-call training. Devstral is the default; Qwen-class coders (30B where 24GB+ memory is available, 7B fallback) are the common picks. Rule of thumb: the largest tool-calling model that stays responsive on your machine.

Q

Are local agents as good as cloud frontier models?

For everyday edits on a private codebase, a good 30B-class local model lands most of the value at zero marginal cost. For the hardest multi-file refactors, frontier cloud models still lead — which is why Anvaya supports both and lets you escalate per task instead of paying per token for everything.

Q

How is this different from Cline, OpenCode, or Aider with Ollama?

Those are excellent harnesses you can point at Ollama — Anvaya adds the missing layer: a persistent memory graph that compounds across sessions, plus repomap PageRank retrieval and pooled LSP diagnostics. Local inference solves privacy; memory solves the re-learning tax.

Stop Starting From Zero.

One binary. 11+9 Rust crates. 545 tests. Hand-written HNSW index. Three transport modes. Four providers, Ollama, Anthropic, OpenAI, Siemens. Zero API keys required to start. Mind remembers everything after the first session.

Get Started →Read the FAQ