Blog / coding agents
Directionally Local vs OpenCode: Sandboxes and Multi-LLM
OpenCode is a polished open-source coding agent; Directionally Local adds isolated sandboxes and key isolation. How they compare, and how to combine them.
If you're looking for an Aider alternative that supports many LLM providers, OpenCode and Directionally Local both come up. They overlap less than it seems: OpenCode is an excellent open-source coding agent with its own interface, while Directionally Local is an agent plus a sandbox layer built to keep keys and spend under control. You can also run OpenCode inside Directionally Local's sandboxes. Last reviewed October 2026.
What each tool is
OpenCode is an MIT-licensed coding agent with a terminal UI, a desktop app (beta) and editor integration. It supports 75+ LLM providers through the AI SDK and Models.dev, plus local models. It ships a full-access build agent and a read-only plan agent, and a permission system where each tool can be set to allow, ask or deny. It has no built-in sandbox; isolation comes from third-party plugins or from running it somewhere isolated.
Directionally Local pairs Autopilot, an MIT-licensed terminal coding agent and SDK, with Hotcell, an Apache-2.0 sandbox runtime for your own hardware. Autopilot runs on your own key with 25+ providers, OpenRouter by default, and shows the cost of every turn. Hotcell runs Docker, Firecracker or Apple Virtualization sandboxes and keeps real API keys on the host.
Comparison
| Directionally Local | OpenCode | |
|---|---|---|
| Licence | MIT (Autopilot), Apache-2.0 (Hotcell) | MIT |
| Interface | Terminal agent, headless SDK | Terminal UI, desktop app (beta), IDE integration |
| Models | 25+ providers, including OpenRouter, Anthropic, OpenAI and local Ollama or LM Studio, plus any OpenAI-compatible endpoint | 75+ providers via Models.dev, plus local models |
| Isolated sandboxes | Built in: Docker, Firecracker, Apple Virtualization | None built in; third-party plugins |
| API key protection | Real keys stay on the host; sandboxes get revocable per-sandbox tokens | Keys live in your local environment or config |
| Spend controls | Per-sandbox spend caps; cost per turn (billed on OpenRouter, estimated elsewhere) and /cost totals |
Not built in |
| Permissions | Plan, edit and auto modes | Per-tool allow / ask / deny rules |
| Parallel work | Read-only subagents in parallel sandboxes (on OpenRouter); many sandboxes per host | Multiple sessions |
| Privacy | Runs on your machines | Says it doesn't store your code or context; calls go directly to your provider |
Sources: OpenCode README, OpenCode providers, OpenCode permissions, OpenCode enterprise, Hotcell README.
Use them together: OpenCode inside Hotcell
Hotcell can create sandboxes with OpenCode preinstalled and pointed at its key gateway:
npm i -g hotcell
hotcell keys add openrouter
hotcell create -n 5 --name feat --branch auto --opencode --repo https://github.com/you/app
hotcell terminal <id> # inside: cd app && opencode
Each of the five sandboxes gets the repo on its own branch, OpenCode ready to go, and only a short-lived token instead of your real OpenRouter and GitHub keys. git push still works, because each sandbox's git remote goes through the gateway. When you're done, hotcell rm --all removes them all and leaves your repo untouched.
When to pick each
Pick OpenCode on its own if you want a polished interface on your laptop, broad provider support out of the box, and you're comfortable with the agent running directly on your machine with your keys.
Pick Directionally Local if your agents handle secrets, you need enforceable spend caps, or you want isolated sandboxes on your own hardware. Use Autopilot as the agent, or keep OpenCode and run it inside Hotcell.
FAQs
Does OpenCode have a sandbox?
Not built in. Its permission rules can make it ask before editing files or running commands, and community plugins add sandboxing, but by default it runs on your machine.
Which supports more models?
OpenCode supports 75+ providers through Models.dev. Autopilot supports 25+ providers, with native adapters where it matters (such as Anthropic's Messages API), local models through Ollama and LM Studio, and any OpenAI-compatible endpoint for the rest. Both cover the major models; OpenCode's built-in list is longer.
Can I run several agents in parallel?
Yes. Hotcell can run many isolated sandboxes on one host, each with its own branch, and, on OpenRouter, Autopilot runs read-only research subagents in parallel sandboxes. OpenCode supports multiple sessions, without isolation between them.
