Comparison
A self-hosted AI coding agent — entirely in your cloud.
"Self-hosted" gets used loosely. The real question is architectural: does the whole system — control plane included — run on infrastructure you own, or does part of it always phone home to someone else's cloud? Here is the honest, sourced breakdown.
Facts verified as of October 2, 2026, from each vendor's own primary sources (linked below). Re-checked immediately before this page ships — vendor terms and tiers change.
Fully in your cloud, or hybrid?
Shipwright's control plane, agent runtime, and your code all run on infrastructure you own, at every tier — no managed-service layer in between. The rest of the field is either hybrid (part of the system stays in the vendor's cloud) or gates full self-hosting behind a sales conversation.
| Tool | License | Architecture |
|---|---|---|
| Shipwright | MIT | Fully in your own cloud — control plane included. The admin, metrics, and agent runtime all deploy to infrastructure you own (Docker or your own Kubernetes cluster). No managed service sits between your repository and the agent. |
| Cursor (self-hosted cloud agents) | Proprietary | Hybrid — no documented customer-controlled control plane. Self-hosted machines keep your codebase, build outputs and secrets on your own infrastructure, across eight supported compute providers, and the agent runs tool calls there. Cursor documents no option to run the orchestration layer yourself. No spec phase, no deploy stage. [source] |
| OpenHands | MIT root, time-limited PolyForm carve-out on enterprise/ | Productized VPC self-host is sales-gated. The MIT license applies to the local tier; VPC self-hosting for the delivery-pipeline features is sold, not shipped free — the vendor's own site states "Talk to sales" for it. [source] |
| Coder Agents | Governance platform, not an OSS agent | No built-in cloud runtime to compare — it wraps agents you run. Scope stops at code, tests, and PRs. No spec phase, no deploy stage, no built-in cron. [source] |
| Devin | Proprietary | Cognition-hosted control plane at every tier. Devin Outposts can run the execution sandbox in your own environment, but the agent loop — the brain — always resides in Cognition's cloud. Full breakdown at /vs/devin. [source] [source] |
| Factory | Proprietary | Cloud-orchestrated by default — running it yourself is sales-gated. Cloud-Managed and Hybrid both keep orchestration in Factory's cloud (Hybrid limits it to metadata). Factory Private, launched September 2026, does put the control plane in your own VPC or on-premises environment, with air-gapped configurations available — so it genuinely removes Factory's servers from the loop at runtime. It is in private preview and reachable only by contacting their team, with no self-serve path. [source] [source] [source] |
"Self-hosted" is table-stakes marketing language across the field now — the architecture underneath it is what actually differs. Full head to head on Devin specifically: /vs/devin →.
Economics
Shipwright: free, MIT, no tiers or seats — the whole pipeline, not a gated tier of it. Everyone else in the table above: see each vendor's own site for current terms, which change over time.
Try it
Install the plugin into Claude Code and run a task end to end. Free, MIT, runs on your own infra.
/plugin install shipwright@app-vitals/shipwrightRolling this out to a team? We can talk through which stage to start at.How the stages work