Self-hosted AI coding assistants — 2026 buyer's guide.

If you work in finance, healthcare, defense, or any organization where "the data must not leave the network" is a hard constraint, the SaaS AI coding assistants don't fit. This guide is for you. It covers what to look for, the real options in 2026, and the trade-offs nobody puts on a feature page.

What "self-hosted" actually means

There are three different things vendors call self-hosted. Be specific:

• Self-hosted control plane. The vendor's backend runs on your infra. Your code still hits their software but never leaves your network. Common for SOC 2 / compliance fits.
• Self-hosted model. The LLM runs on your hardware (an internal GPU cluster or a desktop). The agent app might still be a desktop client, but no inference happens off-prem.
• Air-gapped. No network connections at all. Install offline, model on local hardware, updates via media transfer.

These compose. The strongest posture — air-gapped, self-hosted model, local agent app — is the right answer for the most sensitive environments. It's also genuinely achievable in 2026, which it wasn't two years ago.

The checklist

• No telemetry. Not "configurable telemetry." None. Verify with tcpdump.
• No mandatory account. If sign-in is required, sign-in goes somewhere — and that somewhere is on the internet.
• Local model support. The app should treat LM Studio / Ollama / your internal inference cluster as first-class providers.
• Signed updates. If you let it auto-update, the binary must verify a vendor signature. If you can't permit auto-update, manual update via a verified installer.
• Documented network calls. Every connection the app makes, listed in the docs, with a way to disable each.
• Approval gates. Reads auto-execute; writes pause for a human. Non-negotiable.
• Audit log export. For regulated environments, every tool call must be logable in a format your SIEM understands.
• Identity integration. SAML/OIDC for sign-in (where applicable), SCIM for provisioning, MDM-friendly install.
• Source disclosure. Either open source, or the installers and updater manifests are public — so you can verify what you're shipping to your fleet.
• Procurement-friendly licensing. Site licenses, named accounts, paper contracts available.

The real options in 2026

AZMX AI

Native 7 MB desktop app. BYOK across every major provider or fully offline via LM Studio / Ollama. No account. No telemetry. One signed update check the user can block. Approval gates by default. Deny-list refuses .env, .ssh, credentials. Installers and the updater manifest are public; the Enterprise / Gov tier adds self-hosted, FIPS, PIV/CAC, SOC 2, SIEM export, and named SLA. The right starting point for most regulated teams.

Continue (open source)

VS Code / JetBrains extension that points at any model gateway. Strong fit if you already run an internal LLM gateway and want to plug a coding agent into it. More assembly required.

Aider with a local model

Terminal-native, open source, points at any provider including local. The minimalist's choice — no GUI, no extension store, just the agent and your shell.

Tabby

Self-hosted code completion + chat. Open source. Strong on completion, lighter on agentic tool use; if your need is more "Copilot that doesn't phone home" than "full agent," Tabby is the right shape.

Sourcegraph Cody (Enterprise)

Self-hosted control plane with deep code search integration. The most enterprise-grade procurement story in the category, with proportional pricing.

Windsurf (Enterprise)

Self-hosted control plane for the Cascade agent. Strong enterprise distribution story; full IDE experience. Heavier than the alternatives.

The trade-offs nobody puts on a feature page

Local model quality lags. The frontier models still have an edge on the hardest tasks. The right pattern is hybrid: local model for the 80%, BYOK frontier (where compliance allows) for the rest. AZMX AI is built for this.

Self-hosted ops is real ops. If you choose "self-hosted control plane," budget for it. Someone now operates an LLM stack. Most teams underestimate the GPU cost and the on-call.

Procurement loves a paper trail. Pick a vendor that can sign a real contract, name a real SLA, and produce a real architecture diagram. Open source alone isn't sufficient for many compliance programs.

How to choose, in three questions

1. Can data leave the device? If no — air-gapped, local model. AZMX AI + LM Studio / Ollama is the path of least resistance.
2. Can data leave the network? If no — self-hosted model on your infra, agent app on developer machines. AZMX AI pointed at an internal inference cluster works today.
3. Can data leave the org? If yes — BYOK to your enterprise contract with OpenAI/Anthropic/Google. AZMX AI handles this without changing app.

Talk to us about Enterprise / Gov · Download AZMX AI