Delv
CLIby OpenCode★ 4.3

OpenCode

Open-source terminal coding agent that ships with native MCP support. Self-hostable, BYO-model. The community alternative to Claude Code.

Homepage Source
B
Safety & Trust

Delv Safety Grade: B

Score 72/100 · assessed 2026-04-19

Maintainer55
Permissions40
Supply chain75
Transparency85
Incidents100

OpenCode is an open-source terminal coding agent with native MCP support, positioning itself as a community alternative to Claude Code. The project shows good transparency with active GitHub presence and clear open-source licensing. Supply chain is reasonable via standard package distribution, though the project is relatively new. The main safety concern is the inherent permissions model: as a coding agent, it requires shell execution and filesystem write access to function, which creates meaningful attack surface. The maintainer appears to be a smaller team or solo developer rather than an established organisation, which introduces bus factor risk. No known security incidents, and the self-hosted BYO-model approach gives users control over the LLM layer. MCP integration is partial (tools and resources only), which slightly limits exposure compared to full-featured implementations.

Green flags

  • Fully open-source with active GitHub repository and transparent development
  • Self-hosted BYO-model approach gives users control over LLM selection
  • Standard package distribution via npm/pypi reduces supply chain opacity
  • Clear config at ~/.opencode/config.json makes MCP setup auditable
  • No known security incidents or malicious versions

Red flags

  • Requires shell:execute and filesystem:write for core coding functionality
  • Relatively new project with unclear long-term maintenance commitment
  • Smaller maintainer team increases bus factor and review capacity concerns
  • No prompts capability limits ability to audit pre-built workflows from servers

Permissions requested

Shell executeRead filesWrite filesDelete filesOutbound networkRead envExternal LLM call
Assessed by Delv Editorial using public metadata. Grades are advisory and update as the ecosystem changes. They do not replace your own review of permissions and code before granting an agent access to sensitive systems.

MCP capabilities

  • Tools
  • Resources
  • Prompts
  • Sampling

Platforms

macOSLinuxWindows

Config location

~/.opencode/config.json

Review

OpenCode is what happens when someone looks at Claude Code and decides the world needs an open alternative that doesn't lock you into Anthropic's ecosystem. It's a terminal coding agent that treats MCP as a first-class citizen, not an afterthought. You bring your own model, host it yourself, and wire in whatever MCP servers you need through a straightforward JSON config at ~/.opencode/config.json. The MCP support is solid where it counts. Tools and resources work as expected, which covers most practical use cases. No prompts capability means you can't expose pre-built workflows from servers, and no sampling means the agent can't delegate sub-tasks back to the LLM through MCP. For day-to-day coding work, I haven't missed either. I've been running it against a local Qwen instance and a handful of MCP servers: filesystem access, a Postgres inspector, and a custom tool for our internal API docs. The setup took maybe ten minutes. The config format is clean, no nested YAML nightmares. You point it at your model endpoint, list your MCP servers with their command invocations, and you're done. Where it shines is control. You're not at the mercy of Claude's context window pricing or rate limits. You pick the model, you own the data, you decide what runs where. For teams with compliance requirements or anyone who just wants to experiment with different models, that's the entire point. The rough edges are what you'd expect from a community project. Error messages can be cryptic when an MCP server misbehaves. The UI is functional but spartan compared to polished commercial tools. Documentation exists but assumes you've read the MCP spec. If you need hand-holding, this isn't it. Performance depends entirely on your model choice. With a capable local LLM, it's fast. With a slow one, it's painful. That's on you, not OpenCode. I reach for this when I'm prototyping MCP integrations or working on projects where I can't send code to third-party APIs. It's not trying to be Claude Code. It's trying to be the version you can audit, fork, and run on your own terms. For that specific job, it succeeds.
Verdict

If you want MCP support without vendor lock-in and you're comfortable self-hosting, OpenCode delivers. If you need polish, prompts support, or don't want to manage your own models, stick with Claude Desktop.

Good at

  • Genuine model freedom, works with any OpenAI-compatible endpoint
  • Clean MCP config, no proprietary formats or hidden complexity
  • Self-hostable with full control over data and execution
  • Active community development, forkable codebase
  • Cross-platform CLI that fits terminal-first workflows

Watch out

  • No prompts or sampling MCP capabilities
  • Documentation assumes MCP familiarity
  • Error handling can be opaque when servers fail
  • UI is minimal, no visual niceties
  • Performance entirely dependent on your model choice