#OpenCodeGo

image generation has become an important tool for coding agents you can use it to generate concept UI or reference art for video games OpenCode 2.2.3 has an image gen tool that supports image models from OpenAI, Gemini, Grok and more

OpenCode Go is now available to your entire team

Operation Cheepseek Phase 4: $100 of usage for DeepSeek V4.1 Flash

GET 30M TOKENS FOR FREE you can get access to Qwen3.8-27B and also get 30M tokens every month works with OpenCode, Cline, Roo Code, Continue, and your own API projects. it's very easy just sign up, create an API key and set up in OpenCode or other AI agents. try it here:

🛎️ Round 3: DeepSeek API vs OpenCode Go vs OpenDesign vs Ollama. Now in the Pi harness! Who do you think did better? Rune Compiler is a puzzle game where players connect runes to build small programs. It tests whether a model can: - Build a working compiler, simulation, and 3D interface - Follow detailed rules across eight puzzles - Use automated tests to find and fix errors Ollama 25:39 wall time, 3.84s mean TTFT, 7.58s p95, 14.39M tokens, 0 inference errors DeepSeek API 44:06 wall time, 4.24s mean TTFT, 7.58s p95, 24.75M tokens, 0 inference errors OpenCode Go 70:19 wall time, 8.08s mean TTFT, 7.57s p95, 28.14M tokens, 0 inference errors Two first-token waits of 195s and 256s raised the mean OpenDesign 119:37 wall time, 5.76s mean TTFT, 9.88s p95, 31.63M tokens, 3 inference errors Two connection errors and one billing-scope timeout Demos are available here:

🔔 Round 2 is COMPLETE: DeepSeek API vs. OpenCode Go vs. OpenDesign vs. Ollama. Who do you think won? The challenge: Cartographer’s Paradox, a logic-heavy hex-map game where you use clues from an unreliable cartographer to uncover terrain and identify the lies. This test probes: - Reasoning: Translating complex rules into game logic, including seeded puzzle generation. - Design and engineering: Building procedural 3D graphics and interactive controls, then testing the result against a detailed prompt. Here are the demos:

/goal is finally in @opencode enjoy

pretty cool how they did it. It uses the Claude agents sdk, which calls the claude binary, but with the sdk then turn off using claude's features and instead route it back to opencode for stuff like tool use. This is just a dumbed down tldr, but i'd be curious if someone knows how customizable agents sdk let's you go, and if there were any tradeoffs (still needing to use claude features in some areas) with this design, maybe @thdxr ? Will also be cool to see if @pidotdev and other third party harnesses adopt this

you can now use your claude subscription in OpenCode

no matter what agent you're using, your voice is the best way to talk to them and hex is the best way to use your voice: - fast - runs locally - works on mac, windows, linux stable release now available:

ANTHROPIC ACABA DE MATAR EL “AI SLOP” EN DISEÑO WEB Se llama Frontend Design. • 185.000 estrellas en GitHub. • +760.000 instalaciones. • Open source y mantenida por Anthropic. El problema que resuelve es común: Le pedías una web a la IA y siempre salía lo mismo. → Inter como fuente → gradiente morado → layout genérico de startup → el mismo hero de siempre Frontend Design lo corta de raíz. Obliga al modelo a comprometerse con una dirección estética ANTES de escribir una sola línea de código. y ahí cambia todo. le puedes pedir: → brutalista → editorial → retro-futurista → lujoso → maximalista → orgánico Y te devuelve HTML, CSS, JS, React o Vue listos para producción. Tipografía, color con variables CSS, motion y composición coherentes entre sí. Cómo instalarla: 1. npx skills add --skill frontend-design 2. Describe la web o componente 3. Elige la dirección estética 4. Claude monta el sistema visual completo Una sola instalación. Disponible en todas tus sesiones. Funciona en Claude Code, Codex, Cursor, Gemini CLI, OpenCode y más. +760.000 instalaciones dicen más que cualquier review.

i’ve been looking more closely at the new OpenCode Go DeepSeek setup. 5 things stood out: 1. $100 of “usage” for $10 sounds insane, but i care more about completed tasks than headline credits. if the model needs 3x more turns and tool calls, the math changes fast. 2. the biggest red flag for me is context quality. i’ve seen cases where Go starts losing track around 200K context, while DeepSeek is usually much better than that on long sessions. 3. speed is clearly part of the tradeoff. around 50–60 tok/s plus extra latency between turns is fine for background agents, but annoying for interactive work. 4. parallel sessions can burn allowance faster than the raw token count suggests. that makes the effective value harder to predict if i’m running several coding agents at once. 5. i’m increasingly separating worker and orchestrator roles. DeepSeek on Go makes sense for cheap implementation, but i’d still rather use Opus for planning, review and keeping a complex project on track. my takeaway: OpenCode Go is still great value. but “$100 for $10” matters a lot less to me than cost per task that actually finishes correctly.

OpenCode just got a big update. You can use your Claude subscription in it again. What else came with it: > The Go plan costs $10 a month, and its DeepSeek V4.1 Flash limit went up to $100 a month > A built-in image generation tool that supports OpenAI, Gemini, Grok and other providers So one open source terminal agent can now run on your Claude subscription, a cheap DeepSeek plan or whichever image model you prefer.

Announcing plugins for OpenCode Desktop Make your own

Oh here’s my opencode prompt completion

i grabbed all the community GUI plugins for OpenCode Desktop so far. here are some of them: github inside desktop: commit graph, PRs, reviews, CI, issues (by @Neriousy) T3 Code style PR review with inline comments and "review with agent" a notepad + checklist per chat that the agent can add to (by @parzival1l) codex style side chats that branch off your current chat a plugin store inside settings to browse + install plugins turn plugins on/off and insert @skills from the session header offline voice prompts with a local model a global kanban board (by @dbpolito) /schedule 1h <prompt> for deferred prompts cron jobs on macOS launchd with a desktop panel an animated eye that reacts to what the agent is doing codex pets in desktop live tok/s, time to first token, cost and run history live tok/s badge in the session header go usage ring in the tab bar ghost text autocomplete in the composer, tab to accept search inside a session with highlighting column filters on markdown tables in chat a rail that jumps between your prompts a live markdown + mermaid doc per session handoff briefs you can review before sending (by @mustafaerpek) terminal multiplexer panes you can type into model details panel + switcher, plus git/coffee/budget pills (by @R_Sallusti) environment workbench: status line, PR pill, claude skin review, fix, re-review loop with copilot drive a real chrome/edge browser with a preview computer use for windows + android (by @SMTECHYT) live previews of single-file TSX pages with PDF export translate any selected text RTL topbar fix for windows 8 small ones: keep-going, bubble-tabs, quick-quote, pr-link, blindfold and more and a starter template if you want to make your own (by @alspinu) keep shipping 🫡

we get occasional reports that deepseek in OpenCode Go feels dumber for some people i ran all the benchmarks i could find against the different configs we have it's all in the same range, we did discover 2 specific issues in some low traffic configs we're going to kill

Polls are in. I hear you all loud and clear. BOTH it is! This experiment continues is gonna be a royal rumble between DeepSeek API vs OpenCode Go vs OpenDesign vs Ollama. Why? A follower generously offered his OpenDesign API key so that I can use it to test. That allowed me to purchase an Ollama subscription for the month. Our experiments continue!

@claudeai subscriptions in @opencode are back! Check it out

FREE OPEN-SOURCE GUI FOR CLAUDE CODE + CODEX + CURSOR 😳 it's called T3 Code built by @theo one window that runs the coding agents you already have plus free iOS and Android apps so you can drive them from your phone still early (expect bugs) and your computer has to stay on for phone remote good for: - Claude Code - Codex - Cursor - Grok Build - OpenCode - Antigravity you don't pay for T3 itself just use the plans you already bought how: > install one of those agent CLIs and log in > grab T3 Code (link in the reply) > open your project and pick an agent > scan the qr with the phone app for remote which agent would you throw in first?

since OpenCode has access to its own api, it can query history of the current session combined with codemode it can do a targeted search that doesn't bloat context here it's looking up how it dealt with uber eats in the past, something it lost in compaction

A GITHUB PROJECT THAT HELPS YOU KEEP CODING WHEN YOUR AI MODEL RUNS OUT OF QUOTA. It’s called OmniRoute. It connects to 375 AI providers through a single endpoint, with over 150 reportedly offering free tiers. Here’s what makes it interesting: → Automatically switches to another model when your current one hits its quota → Works with Claude Code, Codex, Cursor, OpenCode, Cline, and Copilot → Compresses conversation context to help reduce token usage → Claims token savings of 15% to 95%, depending on the workload → Runs locally on your device → Open source under the MIT license Instead of manually switching between providers when you hit a limit, OmniRoute helps route your requests to available models. If you’re constantly coding with AI and running into usage limits, this is worth checking out. 79.1K GitHub stars. Repository link below.

How to run an AI coding agent for $0 instead of $200 a month (FULL GUIDE) A GitHub repo with 170k+ stars OpenCode is an open-source coding agent for your terminal, MIT license It reads your project, edits files and runs commands You choose the model behind it, and 75+ providers are supported Repo: Here's how to set it up in 10 minutes 1. How it works OpenCode is only the agent. The model is your choice, and there are 3 free ways to get one: • Free models built into OpenCode, no account needed • Free API keys: NVIDIA NIM, OpenRouter • Ollama: cloud models on a free tier, or local models on your machine 2. Install Mac / Linux: • curl -fsSL | bash Windows: • npm install -g opencode-ai Go to your project folder and run: • opencode 3. Easiest start: built-in free models OpenCode ships with free models that need no signup and no API key • Type /models • Pick any model marked Free (for example Big Pickle, DeepSeek V4 Flash Free, Nemotron 3 Ultra Free) • Send your first message The list changes often, so /models always shows what's available right now Limits are shared, so if one model says rate limit, switch to another 4. Free NVIDIA NIM key • Get a free key at (no credit card) • In OpenCode run /connect, choose NVIDIA and paste the key • Run /models and pick a model. Start with a Nemotron one The limit is about 40 requests per minute. An agent makes many small requests, so long tasks can hit it 5. OpenRouter free models • Create a key at • Run /connect, choose OpenRouter, paste the key • In /models pick any model with :free in the name Limits: 20 requests per minute and 50 requests per day. If you buy $10 of credits once, the daily limit goes up to 1,000. One agent task can easily burn 50 requests, so keep this as a backup 6. Ollama: cloud or local Install Ollama (version 0.15 or newer) Cloud, no GPU needed: • ollama signin • ollama launch opencode --model glm-4.7:cloud The free tier has usage limits per session and per week Local, runs on your machine: • ollama pull qwen3-coder:30b • ollama launch opencode Ollama connects OpenCode by itself. A 30B model needs about 24GB of VRAM or unified memory On less, pick a smaller model or use the cloud one 7. Fix the context length OpenCode needs at least 64k tokens of context. On many setups Ollama starts with 4,096 When the window overflows, the agent forgets files and tool calls break, and it looks like the model is bad Start the server with a bigger window: • OLLAMA_CONTEXT_LENGTH=65536 ollama serve If you use the Ollama app, change the context length in its settings 8. What to pick • Just want to try it → built-in free models • Need more requests → NVIDIA NIM • Weak laptop, long sessions → Ollama cloud • Private code → local Ollama only 9. Read this before you start • Free models are weaker than Claude or GPT on hard tasks. Use them for boilerplate, tests, refactors and explaining code • Built-in free models are offered for a limited time and the teams use them to collect feedback. Don't send private code there • Free tiers and model lists change fast. Check the current limits before you rely on one You don't need a subscription to start coding with an agent You need a terminal and one model that works Like/RT/Bookmark if this was helpful!