opencode — Free, open-source AI coding agent for your terminal
Stop paying monthly for an AI wrapper. opencode runs in your terminal, connects to 75+ model providers via your own API keys, and costs exactly what you use — nothing more.
I switched from Cursor after my third renewal. The thing that finally pushed me: I realized I was paying for the tool and the model, when I already had API credits sitting unused. opencode let me plug those in directly.
A few things that actually matter in daily use:
Build vs Plan mode. Plan mode drafts what it's going to do before touching any files. Sounds small. Isn't.
Provider switching on the fly. Heavy reasoning on Sonnet, routine cleanup on something cheaper. No config change, no restart.
It's just a terminal app. No Electron, no background process, no account required to start.
One honest note: the Claude Pro/Max subscription login was removed in v1.3.0 after Anthropic pushed back. API key only now. For me that was already the plan anyway.
if u are interested with this ， try to quick start opencode
Replies
Is there a learning curve in the beginning?
The API-key-only model is actually a feature for production use — it forces you to own your costs and usage visibility rather than abstracting them away. The 75+ provider support is the practical differentiator over tools like Claude Code or Cursor where you’re locked into one provider’s pricing and rate limits. The interesting question is context management: terminal-native agents tend to struggle with large repository context because feeding everything into the prompt is expensive and noisy. How does opencode handle context windowing for large codebases — file-level summaries, embeddings-based retrieval, or raw inclusion?
@jimmypk Interesting question — I did some digging and compared how OpenCode and Claude Code agents handle context differently:
Dimension
OpenCode
Claude Code
Raw Inclusion
❌ Never
❌ Never
Embedding-based RAG
❌ Not implemented
❌ Not implemented
LSP Symbol Indexing
✅ Native, first-class
⚠️ Limited
On-demand Tool Retrieval
✅ Core mechanism
✅ Core mechanism
File-level Summarization
⚠️ Implicit compression
✅ Explicit SUMMARIZE_TOOL_RESULTS_SECTION
Static/Dynamic Prompt Split
❌
✅ Prompt cache optimization
Sub-agent Delegation
❌ Single-agent only
✅ EXPLORE_SUBAGENT parallel exploration
Cross-session Memory
❌ Stateless
✅ Agent Memory persistence
Quick takeaways:
Neither of them does raw inclusion or RAG — both rely on lazy, tool-driven retrieval, only pulling in what the agent actually needs
OpenCode's advantage: LSP handles symbol-level navigation without burning through tokens — pretty efficient
Claude Code's advantage: sub-agents spin up parallel explorations each with a fresh context window, which is a big deal on large monorepos
Where both fall short though — semantic search like "find all payment failure handlers" — neither comes close to Cursor's embedding pipeline for that kind of fuzzy cross-module reasoning
Hello Aria
Terminal-native coding agent that’s free and open source? That’s going to become a default tool for a lot of developers who don’t want to be locked into an IDE plugin ecosystem.
The open-source angle is particularly meaningful right now — there’s a real wave of developers who want to inspect what’s happening under the hood when an AI touches their code. We’ve seen a similar "trust through transparency" dynamic in the AI personal assistant space with Hello Aria (our day management AI on WhatsApp/iOS) — users want to understand what the AI is doing on their behalf, not just trust a black box.
Opencode in the terminal is just the right UX for devs who live there. Would love to see integrations with tmux layouts down the line. What LLMs are supported out of the box?
Do you miss anything important from cursor after switching?