A self-hosted stack for AI-assisted development
Which parts of an agent workflow can run on your own infrastructure, which cannot, and what each choice costs in quality and effort.
Most of an agent workflow can run on your own machines. The model is the part that is genuinely hard to bring in-house, and everything else is a choice. Being clear about which is which keeps the project from stalling on the wrong problem.
The components
| Layer | Self-hostable | Notes |
|---|---|---|
| Repository | Yes | Git on any internal host, or a bare repo over SSH |
| Semantic index and diff | Yes | Local computation over your checkout |
| Agent CLI | Yes | Runs where you install it, points where you tell it |
| Model | Partly | Open weights on your own GPUs, at a quality cost |
| Team sync and messaging | Yes | A service you run, with your own database |
| Frontier model quality | No | The honest limit |
The model decision
There are three positions, and choosing between them is a policy question rather than a technical one. Use a hosted frontier model and accept that prompts leave your network under a contract. Serve open weights locally, accept lower capability, and compensate with tighter verification. Or split: local models for bulk mechanical work, a hosted model for the hard parts, with a rule about which code may go where.
What to run on your own hardware
- An OpenAI-compatible inference server, so agent CLIs need only a base URL change.
- A package mirror, or vendored dependencies, so builds do not depend on a public registry.
- The Git host, with backups you have restored at least once.
- The coordination service, if you run several people or agents on shared work.
What it costs to operate
GPUs, someone to keep the inference server running, and a mirror that needs updating. That is a real operational load, and the honest comparison is against per-seat subscriptions for the whole team plus the policy risk of code leaving the network. For some organisations the arithmetic is obvious in one direction and for others in the other.
Where Aura fits
The engine, the CLI and the desktop app are Apache 2.0 and run locally with no account. The collaboration server is a service you can run yourself, so team sync, messaging and the shared history stay inside your network. The only outbound traffic is whatever your chosen model provider needs, which is a decision you keep rather than one the tool makes.
See also: running with no internet at all and open-source alternatives to hosted agents.
Can you self-host an AI coding stack?
Most of it. Version control, the agent harness, indexing, review tooling and the audit record all run on your own hardware. The frontier models do not — that is the one component you either replace with local weights or accept as an external dependency.
What are the parts of a self-hosted AI coding setup?
A model runtime, an agent harness that drives it, a repository host, a package mirror, and somewhere to keep the record of what the agents did. Each can be swapped independently; the harness is where most of the behaviour you care about lives.
Why self-host AI coding tools?
Code never leaves your network, the tooling does not change under you, and you keep working when a vendor has an outage or a policy change. The cost is that you are now operating it, and that local models are behind on the hardest tasks.
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.