OpenHands

also known as OpenDevin

Type: full-code · Vendor: All Hands AI · Language: Python · License: MIT (enterprise/ excluded) · Status: active · Status in practice: mature · First released: 2024-03-13

Links: homepage docs repo

Open-source AI software-development platform: a composable Python SDK plus CLI, local GUI, and cloud GUI that runs tool-using agents inside an isolated sandbox (Docker / process / remote) with MCP tool extension and AGENTS.md-based skills.

Description. OpenHands (formerly OpenDevin) is a community project for AI-driven development. The core is a composable Python SDK that runs an agent loop: an LLM-driven agent executes bash, edits files, browses the web, and reaches external tools via MCP. Execution happens inside a configurable sandbox (Docker by default, process or remote alternatives), which gives strong host isolation. Repository-wide instructions live in AGENTS.md (or CLAUDE.md / GEMINI.md); on-demand skills live under .agents/skills/ or the legacy .openhands/microagents/ directory. OpenHands ships as CLI, Local GUI, and OpenHands Cloud on the same SDK.

Agent loop shape. Tool-using agent loop: the agent proposes a tool call (bash, file edit, browser, or MCP-provided tool); the call executes inside the configured sandbox; the resulting observation is returned to the model; the loop repeats until the task is done. AGENTS.md content is injected at conversation start; on-demand skills under .agents/skills/ are summarised first and read fully only when needed.

Primary use cases

sandboxed autonomous software-development agent
bash + file-edit + browser tool use inside Docker
MCP-extended toolchain for code agents
AGENTS.md-driven project skills
self-hosted alternative to closed cloud coding agents

flowchart TD user[User task] --> ctx[AGENTS.md + skills context] ctx --> loop[Agent loop] loop --> propose[LLM proposes tool call] propose --> sandbox{Sandbox} sandbox -->|docker| dock[Docker isolation] sandbox -->|process| proc[Local process] sandbox -->|remote| rem[Remote sandbox] dock --> exec[Execute action] proc --> exec rem --> exec exec --> bash[Bash] exec --> edit[File edit] exec --> browser[Browser] exec --> mcp[MCP tool] bash --> obs[Observation] edit --> obs browser --> obs mcp --> obs obs --> loop loop --> done[Done]

Key concepts

Sandbox → sandbox-isolation (docs) — Execution environment where OpenHands runs commands, edits files, and starts servers. Docker / process / remote.
Agent / Tools / Workspace / Conversation (docs) — Core SDK abstractions: an LLM-powered Agent uses Tools inside a Workspace via a Conversation.
MCP → mcp (docs) — Open standard for external tool servers; OpenHands routes tool calls to configured MCP servers.
AGENTS.md and skills → agent-skills (docs) — Always-on AGENTS.md for repo-wide instructions; on-demand skills under .agents/skills/ or legacy .openhands/microagents/.
Multi-interface SDK (docs) — Same agent runs as CLI, Local GUI, or OpenHands Cloud on a shared SDK.

OpenHands

Neighbourhood

Instantiates

Alternatives & relatives

Listed as alternative by (3)

References

Provenance