Claude Code, Codex or OpenCode on a local model (Ollama)

Plenty of people have tried putting Claude Code or OpenCode on top of a local model this year, and the complaint I keep seeing in their threads isn't "the model is dumb", it's "the agent forgets the task after three files." Almost every time, that's a context problem, not a model problem. This guide is the map I wish existed: what each tool is, how it connects to Ollama, and how much GPU memory the setup the docs ask for actually costs.

How this works: three agents, one local server

All three are terminal coding agents. They read your repo, edit files and run commands. Normally each one talks to its maker's cloud API. Ollama makes a local model look like those APIs. It added an Anthropic Messages API in v0.14.0 (January 2026), which is what Claude Code speaks. It also offers OpenAI-compatible endpoints, which Codex and OpenCode use. Then v0.15.0 added ollama launch, which configures and starts Claude Code, Codex, OpenCode or Droid for you.

# Ollama v0.34.4 — one-command setup for each agent
ollama launch claude
ollama launch codex
ollama launch opencode

If you already manage models in Ollama, you're most of the way there; our Ollama model-management guide covers pulling, listing and removing models. The rest of this page is about the two decisions that matter: which agent, and whether your GPU can give it enough context.

The 64k context trap (read this before anything else)

Every one of these integrations asks for a lot of context. The Ollama docs say it plainly for each tool. Claude Code: "set the context length to 64k or higher" for larger repositories. OpenCode: "requires a context length of 64k or higher". Codex: "at least 64k tokens". Ollama's context-length page says coding tools "should be set to at least 64000 tokens."

Ollama's defaults don't give you that. The same page lists the defaults by VRAM: under 24 GiB, 4k context; 24–48 GiB, 32k; 48 GiB and up, 256k. The older FAQ still says a flat 4,096 tokens. Either way, anyone on a normal gaming card starts well short of 64k, and the agent silently loses the early part of the conversation. OpenCode's own provider docs point at the same cause: "If tool calls aren't working, try increasing num_ctx."

# Ollama docs (context-length), accessed 25 Sep 2026
OLLAMA_CONTEXT_LENGTH=64000 ollama serve

# check what you actually got — look at the CONTEXT and PROCESSOR columns
ollama ps

ollama ps is the honest check. If PROCESSOR shows anything other than 100% GPU, part of the model has spilled into system RAM and generation will crawl. That's the next problem: 64k of context costs memory.

What 64k context costs in VRAM

Here's our VRAM engine on the two local coding models Ollama users reach for most. Assumptions: Q4_K_M weights, fp16 KV cache (Ollama's default), 0.8 GiB overhead. The KV cache, the memory that holds your context, grows linearly with context length, which is why 4k looks fine and 64k doesn't.

The lever that helps most isn't a smaller model, it's a smaller KV cache. Ollama's FAQ documents OLLAMA_KV_CACHE_TYPE=q8_0, which uses "approximately 1/2 the memory of f16" with "a very small loss in precision". It needs Flash Attention, which Ollama turns on automatically where the hardware supports it. On our numbers, that takes Qwen3-Coder at 64k from 24.0 to about 21.0 GiB, which is the difference between "tight" and "fine" on a 24GB card. Full per-model tables are on best GPU for Qwen3-Coder and best GPU for gpt-oss 20B.

# Ollama FAQ, accessed 25 Sep 2026 — global setting, applies to every model
OLLAMA_KV_CACHE_TYPE=q8_0 OLLAMA_CONTEXT_LENGTH=64000 ollama serve

Referral link

Your card can't hold a 64k-context coder? Rent one by the hour

On Vast.ai's spot market, 48GB cards like the L40 and RTX A6000 had listings from about $0.34–$0.37/hr (lowest live listings, 25 Sep 2026). That's enough for Qwen3-Coder at 64k with room to spare. Run Ollama on the rented box and point your agent's base URL at it. Spot instances can be reclaimed, so commit your work often.

See live 48GB GPU pricesCompare live prices first →

Referral link — we may earn a commission at no cost to you. Prices on our compare page are the provider's live figures, cheapest first; this never changes the ranking.

Not sure renting beats buying for your hours? Our cost-to-run page for Qwen3-Coder does the per-hour maths from live prices, and can you run Qwen3-Coder locally covers the hardware side in more depth.

Which agent to pick: what the docs actually say

I'm only comparing things you can check in each project's docs or repo: licence, how it connects to Ollama, how it's configured, and what's documented as unsupported. I'm not going to tell you which one "feels smarter". On a local backend they all use the same model, so the model and your context budget matter more than the wrapper.

• OpenCode if you want an open-source tool (MIT) whose config is one JSON file you can commit. Set it up with Ollama →
• Codex CLI if you want Apache-2.0 and plan to use OpenAI's open-weight gpt-oss models; Ollama's docs use them in every Codex example. Set it up with Ollama →
• Claude Code if you already use it with Anthropic's models and want the same workflow offline, and you're fine with a proprietary licence and the API gaps Ollama lists. Set it up with Ollama →

Claude Code vs OpenCode on a local model

This is the matchup people search for most, so here it is head-to-head. The short version: they're closer than you'd think once both are pointed at the same Ollama model. The differences are licensing, configuration and API coverage.

My take: if you're starting fresh on a local model, OpenCode is the lower-friction choice. It's open source, its Ollama setup is one config block, and Ollama doesn't list any API gaps for it. If you already work in Claude Code every day, the Ollama route lets you keep your habits offline. Just know that Ollama's Anthropic compatibility page lists tool_choice and the token-counting endpoint as unsupported, so anything that relies on them will behave differently than against Anthropic's own API.

Already using VS Code?

If you'd rather stay in the editor than live in a terminal, that's a different setup. Our local AI coding assistant for VS Code guide and the Qwen3-Coder in VS Code walkthrough cover the extension route. The same 64k context and VRAM maths above still applies.