Skip to content
Shiba LogoShibaDocs
Source

Shiba vs hosted coding agents

This page compares deployment models, not feature checklists copied from any vendor site. Product names below refer to common categories of hosted coding agents (multi-tenant SaaS with credits, VMs, and managed integrations). Read each vendor’s own docs for their current feature set and pricing.

Dimension Self-hosted Shiba (this project) Typical hosted coding agent
Where it runs Your Cloudflare account (Worker + Durable Objects + Containers) Vendor’s cloud
Source license AGPL-3.0-only — fork and publish changes Proprietary
Model billing Your AI Gateway / provider accounts (BYOK) Often vendor credits, sometimes BYOK with platform fees
Machine billing Cloudflare Containers metered usage Often per-machine-hour credits on top of model cost
Approval before execution Frozen input hash; human approves before any sandbox starts Varies; many products approve per action in-product
Credentials in the sandbox Deny-by-default egress; real keys injected outside the container Varies; reputable vendors also keep git tokens off the VM
Ops burden You provision CF resources, secrets, Access, gateway keys Vendor operates scale, SSO, support tiers
  • You need data and runs to stay in an account you control (single-tenant by architecture, not by contract add-on).
  • You already pay for Cloudflare Workers and want agent compute beside your existing edge stack.
  • You want AGPL assurances for forks and hosted derivatives.
  • You accept operator work: Access, secrets hygiene, gateway BYOK, and honest limits (container size, concurrency caps in wrangler.jsonc).
  • You want managed machine sizes, desktop streaming, or snapshot-heavy VM workflows without operating Containers.
  • You need vendor SOC 2 / enterprise SSO as a packaged offering rather than your own Cloudflare Access + policies.
  • You prefer flat monthly credits over reconciling Cloudflare + provider invoices.
  • Your team will not run deploy tooling (Alchemy or wrangler) or Docker image builds.

Hosted products often emphasize long-lived captains, deep subagent trees, and large VMs with graphical desktops. Shiba today targets an approval-gated pipeline: plan/delegate → sandbox run → optional Worker-side PR publish → optional chained review posted from the Worker (not from the container). Parallelism is flat with configured caps (MAX_CONCURRENT_RUNS, container max_instances). See Architecture and the repository PLAN.md for an honest capability map.

Neither model removes the need to review agent output; Shiba makes the approval step explicit in the product contract (spec/GOAL.md).

  • Shiba: Cloudflare usage (Workers, DO storage, Container vCPU/GiB-seconds) plus your model tokens via AI Gateway. Tokens usually dominate; see Models & costs.
  • Hosted: Subscription or credit bundles plus machine time meters on many public pricing pages. BYOK may still bill machine hours separately.

Numbers change — check Cloudflare and your chosen provider’s current pricing before committing.

  • Shiba: Export is “keep your Cloudflare account”; code is on GitHub under AGPL. Moving away means re-implementing integrations you added (Slack apps, webhooks, custom domains).
  • Hosted: Export is usually API/export features and git history on your repos; agent thread state may not portable.

Do not present this page as an official comparison to any trademarked product. Cite public documentation when you make factual claims elsewhere. Shiba does not ship another vendor’s UI, copy, or assets.