Agentic Coding without Claude and Codex'
Source: Dev.to
Introduction
I started dabbling with agentic coding a few months ago when Claude had decent limits on the $20 plan, and I was really impressed by the “fire and forget” way of doing things. You prompt the agent: I want e2e tests, and it will study the codebase and implement them.
When I began hitting limits on Claude (they reduced limits for everyone on the $20 subscription), I moved to Codex because its limits were more reasonable. After a month with Codex, I started exploring other tools.
OpenCode
OpenCode (https://opencode.ai/) is an OSS agentic coding tool similar to Claude and Codex. It offers features such as:
- Plan mode
- MCP server support
AGENTS.md- SKILLs support
The amazing thing about it is that it supports almost every AI provider out there, and it also supports running local LLMs. One service offered directly by OpenCode is the Zen service (https://opencode.ai/zen), which is essentially pay‑as‑you‑go for curated models.
DeepSeek V4 Pro
I spent a weekend building a side project using DeepSeek V4 Pro with its pay‑as‑you‑go pricing. The model was discounted 75 % at the time; I used around 60 M tokens and paid only $1.87.
I could not sense any difference in quality compared to Claude or Codex. I gave it a vague prompt like “implement end‑to‑end testing with UI,” and it explored the codebase, generated the relevant tests, ran them, iterated, and added them within a few minutes.
Another feature I built with DeepSeek was a page that allows users to:
- Import transactions from a bank statement
- Parse the statement
- Edit transactions
- Confirm the import
This feature was implemented end‑to‑end and polished after a few iterations.
Overall I’m quite satisfied and don’t feel the need to go back to Claude or Codex.
PI.dev
PI.dev (https://pi.dev/) is another OSS agentic coding tool. It’s more minimalistic than OpenCode and lacks MCP support or “Plan Mode.” I haven’t tried it extensively, so I don’t have much to add.
Soulforge
Soulforge (https://github.com/proxysoul/soulforge) takes a different approach. Instead of treating the codebase as plain text, it attempts to build a knowledge graph. I haven’t tried it yet, but I thought it was worth mentioning because it’s an interesting concept.
Conclusion
If you’re unhappy with current providers, you can try other tools and providers. It’s not just Codex and Claude. After coding a whole weekend with DeepSeek V4 Pro and OpenCode, I didn’t notice any significant difference; once I got used to OpenCode, it felt like I was still coding with Codex/Claude.
Thanks for reading!