Running coding agents on an air-gapped network
What it takes to run AI coding agents with no route to the public internet: local models, offline tooling, and how to move artefacts across the boundary without breaking it.
An air-gapped setup means no component may reach the public internet, including the model. That rules out every hosted agent, and it changes which parts of the toolchain you can use, but it does not rule out agent-assisted development.
The four pieces
- The model. Served locally: an open-weights model on your own hardware, behind an OpenAI-compatible endpoint so the agent CLIs can talk to it.
- The agent. A CLI that can point at an arbitrary base URL. Most can; check before committing to one.
- The tooling. Anything that phones home for licence checks, telemetry or model catalogues will fail closed. Test with the network actually off, not with a firewall rule you can remove.
- The repository. Ordinary Git on an internal host.
What breaks first
| Component | Failure | Mitigation |
|---|---|---|
| Package installs | Registry unreachable | Internal mirror, vendored dependencies |
| Model catalogue lookups | Tool cannot list models | A local catalogue file with a pinned fallback |
| Licence or auth checks | Product refuses to start | Choose tools that work without an account |
| Docs and search inside the agent | Tools time out one by one | Disable those tools explicitly rather than letting each fail |
| Telemetry | Slow startup, hung requests | Off by default, and verify |
Moving things across the boundary
Model weights, dependency mirrors and tool updates have to cross once, deliberately, on media that is checked on the way in. The discipline that keeps this workable is to make the crossing rare and reproducible: a bundle with a manifest and hashes, imported by a script, rather than someone copying a directory. Aura's bundle export and import work this way, so a repository with its semantic state and history can move as one verified artefact.
Quality expectations
A locally served open-weights model is behind the frontier hosted ones, and pretending otherwise leads to disappointed teams. What holds up well: mechanical refactors, test writing, migrations, and anything with a tight specification and a fast feedback loop. What holds up badly: open-ended design work. Pair a smaller model with stronger verification rather than expecting the model to be right unaided, which is where an automatic check against the goal earns its place.
What runs offline in Aura
Indexing, semantic diffing, proofs, snapshots, rewind, the intent ledger and the CLI all run locally with no account and no outbound calls. Team sync is opt-in and points at a server you host. See the self-hosted stack for the network-connected version of the same setup.
Can you run AI coding agents without internet access?
Yes, with local model weights and tooling that does not phone home. The agent loop itself needs nothing external. What you give up is the frontier models, which are API-only, and the convenience of fetching packages and documentation on demand.
What do you need to run coding agents air-gapped?
Local weights and a runtime to serve them, a mirror of your package registries, offline documentation, and version control that works without a remote. Then an explicit process for moving artefacts across the boundary, because that becomes the only path in or out.
What quality do you lose using local models for coding?
A real amount on hard, multi-step work — the gap is widest exactly where agents are most useful. On constrained, well-specified tasks with tests to check against, local models are closer than their benchmark scores suggest.
Aura Crew vs Claude Code Loops: the autonomous work-loop, compared
Claude Code defines four loop types — turn-based, goal, time, and proactive. Aura Crew is the same idea productized: a dependency-ordered work-loop with proof, collision-safety, and any agent. Full comparison.
Autonomous coding agent loops, explained: turn-based, goal, time, and proactive
A coding agent loop is an agent repeating cycles of work until a stop condition is met. The four types explained — turn-based, goal-based, time-based, proactive — with when to use each and how to keep quality high.
How to run multiple coding agents in parallel without merge conflicts
Running several AI coding agents at once collides on shared files. The fix: a worktree per agent, a live team radar, and soft-to-hard zone claims so parallel agents never touch the same symbol. How Aura Crew does it.
Goal-based agent loops: giving a coding agent a definition of done
A goal-based loop keeps a coding agent iterating until a success criterion is met. The trick is a deterministic definition of done — tests, a score, a proof — not an LLM guessing at "good enough." How to write one.