Best agent harness for local LLMs and Ollama

The short answer

First fix the context window

Most "the local agent ignores my tools" reports come down to context length. Ollama picks a default from your GPU memory: 4k tokens under 24 GiB of VRAM, 32k from 24 to 48 GiB, 256k at 48 GiB or more. It recommends at least 64,000 tokens for agents and coding tools, and its OpenCode and Codex pages both ask for 64k or more. OpenCode's own tip says that if tool calls aren't working, raise num_ctx, starting around 16k to 32k. Source.

OLLAMA_CONTEXT_LENGTH=64000 ollama serve
# in another terminal: the CONTEXT column shows what was allocated
ollama ps

A larger context needs more memory. If ollama ps shows the model partly on CPU, it will be slow; Ollama advises avoiding CPU offload for best performance.

How each harness connects

Licenses from the GitHub API on Oct 18, 2026. "No license file" means GitHub detects none in anthropics/claude-code.

Rig 1: OpenCode + Ollama

This is OpenCode 1's documented Ollama block, plus a permission rule so edits and shell commands ask first. Replace <model> with a model you've pulled.

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "ollama": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Ollama (local)",
      "options": { "baseURL": "http://localhost:11434/v1" },
      "models": { "<model>": { "name": "<model> (local)" } }
    }
  },
  "permission": { "edit": "ask", "bash": "ask" }
}

Rig 2: Codex CLI + Ollama or LM Studio

# default local provider used when you run with --oss
oss_provider = "ollama"   # or "lmstudio"

codex --oss
# or let Ollama set up a dedicated Codex profile for the session
ollama launch codex

If oss_provider is unset, Codex prompts you to choose. The IDs openai, ollama and lmstudio are reserved, so don't define a custom provider with those names. Config reference.

Rig 3: Aider + Ollama

python -m pip install aider-install && aider-install
export OLLAMA_API_BASE=http://127.0.0.1:11434
ollama pull <model>
OLLAMA_CONTEXT_LENGTH=8192 ollama serve   # Aider's doc example; raise it for bigger repos
aider --model ollama_chat/<model>

Aider recommends the ollama_chat/ prefix over ollama/. Its doc example uses an 8k window; Ollama's own guidance for coding tools is 64k, so start higher if your hardware allows. Aider's repo was last pushed on May 22, 2026, per the GitHub API, which is slower than the other three.

Rig 4: goose + Ollama

goose configure        # Configure Providers -> Ollama
export OLLAMA_HOST=http://localhost:11434
goose session

In goose Desktop, locally installed Ollama models show up in the model dropdown automatically. goose also reads AGENTS.md by default, so the same instructions file works here and in OpenCode and Codex.

Keeping private files local

A local model only keeps data local if nothing else in the rig phones out. Check three things before pointing a harness at private documents:

• Every model slot is local. Harnesses can use more than one model for different tasks; set all of them to your local provider.
• Web tools are off or ask first. In OpenCode, "webfetch": "deny" under permission blocks fetches.
• MCP servers are local. A remote MCP server sends your requests to its host.

See the private-docs recipe for a full read-only rig.

Related