Blog / coding agents
Directionally Local vs OpenHands: Self-Hosted Coding Agents
OpenHands and Directionally Local both run coding agents you can self-host. How they differ on sandboxing, API keys, spend limits and models.
OpenHands and Directionally Local both let you run autonomous coding agents on infrastructure you control. They come at it from different ends: OpenHands is a full agent platform with a web interface and team integrations, while Directionally Local focuses on the agent, the sandbox and keeping credentials away from both. This page compares them honestly so you can pick, or combine, the right one. Last reviewed October 2026.
What each tool is
OpenHands is an MIT-licensed open-source agent platform. You can run it as a CLI, as a local web app (openhands serve, which uses Docker), or through OpenHands Cloud, its commercial hosted offering. Its main repository is now "Agent Canvas", a self-hosted control centre that can run OpenHands and other agents such as Claude Code, Codex, Gemini and OpenCode. By default, the agent runs inside a Docker sandbox. It connects to any model LiteLLM supports, including local models through Ollama, LM Studio or vLLM.
Directionally Local is two open-source tools. Autopilot (MIT) is a terminal coding agent and SDK that runs on your own key with 25+ providers (OpenRouter by default, plus Anthropic, OpenAI, Gemini and local Ollama, among others), with plan, edit and auto modes and the cost shown per turn. Hotcell (Apache-2.0) runs sandboxes on your own hardware as Docker containers, Firecracker microVMs or Apple Virtualization microVMs. Autopilot runs its read-only research subagents in Hotcell sandboxes.
Comparison
| Directionally Local | OpenHands | |
|---|---|---|
| Licence | MIT (Autopilot), Apache-2.0 (Hotcell) | MIT (open-source); Cloud and Enterprise are commercial |
| Interface | Terminal agent, headless SDK | CLI, local web GUI, hosted cloud |
| Sandbox | Docker, Firecracker or Apple Virtualization, on your hardware | Docker sandbox by default; a process mode runs on the host without isolation |
| API keys and secrets | Real keys stay on the host; each sandbox gets a short-lived, revocable token swapped by a gateway | Secrets are exported as environment variables in the agent's runtime |
| Spend limits | Per-sandbox spend caps at the gateway; Autopilot run budgets | LLM budgeting via LiteLLM is listed as an Enterprise feature |
| Outbound network | Optional default-deny egress, kernel-enforced on Linux and microVMs | Depends on your Docker setup |
| Models | 25+ providers, including OpenRouter, Anthropic, OpenAI and local Ollama or LM Studio, plus any OpenAI-compatible endpoint | Any LiteLLM-supported model, including local |
| Team integrations | GitHub via keyless git push from sandboxes |
Slack, GitHub and GitLab in open source; Jira and Bitbucket in Cloud/Enterprise |
| Teams offering | Directionally Local for Teams (early access) | OpenHands Cloud and Enterprise |
Sources: OpenHands sandbox docs, OpenHands secrets docs, OpenHands Enterprise vs OSS, Hotcell README.
The key difference: where your credentials live
In OpenHands, secrets you configure are exported as environment variables inside the agent's runtime. That's practical: the agent can use them directly. It also means anything running in that sandbox, including a malicious dependency or an agent following a prompt injection, can read them.
Hotcell takes the opposite approach. The real API key or GitHub token never enters the sandbox. The sandbox holds a per-sandbox token, and a gateway on the host swaps it for the real credential on the way out, metering the call, enforcing a spend cap and letting you revoke one sandbox without touching the others. The limitation is that this only works for credentials sent in HTTP headers; database passwords and AWS request signing can't be swapped.
When to pick each
Pick OpenHands if you want a web interface, team workflows wired into Slack, GitHub or GitLab, a broad agent platform that can run several agent types, or a hosted option you can move to later.
Pick Directionally Local if your agents handle keys you can't afford to leak, you need hard per-sandbox spend caps, you want microVM isolation on your own Macs or Linux machines, or you're running many agents in parallel on one box.
They aren't mutually exclusive. Hotcell is agent-agnostic, so you can run other agents inside its sandboxes and still keep keys on the host.
FAQs
Is OpenHands fully open source?
The core is MIT-licensed. OpenHands Cloud and its Enterprise features are commercial.
Does Directionally Local send my code to a vendor?
Autopilot and Hotcell run on your machines. Your prompts and code go to whichever model provider you configure; point Autopilot at a self-hosted model if they must stay local too.
Can both use local models?
Yes. OpenHands supports local models through LiteLLM-compatible servers such as Ollama. Autopilot supports Ollama and LM Studio as providers, plus any OpenAI-compatible endpoint you host.
Is there a free way to try Directionally Local?
Yes. Autopilot (npm i -g autopilot-ai) and Hotcell (npm i -g hotcell) are free and open source. Directionally Local for Teams adds a supported deployment for companies.
