Use Case · Offline

Ship Code at 35,000 Feet.

No Wi-Fi, no API, no excuses. The fully offline agent loop — local model, local memory, cached everything.

Direct answer

An offline coding agent runs the entire AI loop without connectivity: local model inference, on-disk memory, cached packages and docs. Anvaya is flight-ready by construction — Ollama models (pull once: devstral default, 7B–30B coders by hardware), the ~15MB binary plus Mind daemon with your project's graph under .anvaya/mind/, and zero license checks or telemetry calls. Pre-cache package registries (Verdaccio/devpi) and a docs mirror (Devdocs/Zeal) before departure; budget 50–80GB disk. What works offline: refactors, debugging, tests, memory recall, the full agent loop. What breaks: uncached installs and post-cutoff library questions. Tested pattern: ship a feature, debug two bugs, write tests — transpacific, zero bars, zero API spend.

Checklist

Flight-Ready in Five.

011 · Model on diskollama pull devstral (or your 7B/30B coder). Verify it answers with networking off — the only test that counts.
022 · Graph travels too.anvaya/mind/ lives in your project tree: decisions, bug threads, and patterns fly with you. Session 200 at altitude recalls session 12 at the office.
033 · Cache the ecosystemRegistry mirrors for your languages, docs mirror, ripgrep-indexed Stack dumps. New-package installs are the #1 offline failure — pre-cache.
044 · Mind the batteryAutocomplete-first sipping beats full-agent-loop gulping on battery. Scope autonomy to plugged-in legs; review diffs on gametime.

Beyond Travel

Offline Is a Superset.

Every offline workflow works online. The reverse fails exactly when it matters.

01Regulated air-gapsRepos that legally can't touch networks get the full agent loop — memory included — where cloud tools are disqualified.
02VPN-only enterprisesInternal registries, staging DBs, and private artifacts stay reachable; nothing routes through vendor clouds.
03Field + unreliable linksDeterministic local latency regardless of bars. Determinism is the feature: same speed, every time.

Questions

Asked About Offline Coding.

Q

What do I need on disk before losing connectivity?

The model (7B ~5GB through 30B ~19GB), Anvaya's ~15MB binary, your project's .anvaya/mind/ graph, cached packages for the languages you'll touch, and a docs mirror (Devdocs/Zeal, a few GB). Budget 50–80GB total and verify with networking disabled before you leave.

Q

What works offline vs what breaks?

Works: autocomplete, refactors, debugging, tests, memory recall, full agent loop. Breaks: installing uncached packages, questions about post-cutoff libraries, web search tools, cloud-provider escalation. Pre-cache dependencies and prefer tasks you could already scope.

Q

Which hardware fits offline agents?

Apple Silicon with 24GB+ unified memory or discrete GPUs with 12GB+ VRAM for 30B-class models; 7B fallbacks run on 8GB machines at ~80–85% quality. Local inference latency (~280ms) beats degraded-network cloud on determinism if not always on speed.

Q

Is offline just for travel?

No — VPN-only enterprise networks, air-gapped regulated repos, field work, and unreliable connectivity share the pattern. Offline-capable is a superset skill: every offline workflow also works online, the reverse isn't true.

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