Guides
Answers to the questions people ask before and after installing a coding agent: running one offline, what it costs, MCP with any model, CI and permissions.
Run an AI coding agent fully offline with Ollama or LM Studio
Point OHarness at Ollama or LM Studio and the whole agent loop runs on your machine: no key, no per-token bill, no code leaving the laptop.
The cheapest way to run a Claude Code-style coding agent
What an agent session actually costs on DeepSeek, GLM, Kimi, Gemini, GPT and Claude — and the free option. One harness, any model, list prices compared.
Use Claude Code with DeepSeek, Ollama or other models
Run Claude Code on DeepSeek, Kimi, GLM or a local Ollama model via ANTHROPIC_BASE_URL: the exact settings, what breaks, and when another agent fits better.
Use MCP servers with GPT, Gemini, DeepSeek or a local model
MCP is a harness feature, not a model feature. Configure a server once in OHarness and every model — GPT, Gemini, DeepSeek, Ollama — can call its tools.
Switch models mid-session without losing context
Move a coding session from Claude to GPT to DeepSeek to a local model with one command. Why most tools restart or convert history, and why OHarness does neither.
Run an AI coding agent headless in CI
oharness-headless runs the agent with no UI and writes JSONL to stdout. Permissions that would prompt are denied, so an unattended run is never less safe than an attended one.
Coding agent permissions: modes, allow and deny rules
An agent runs with your authority. How OHarness decides what runs without asking: five modes, allow/ask/deny rules, where they are stored, and why deny beats bypass.
Use DeepSeek V4 in a coding agent
Run DeepSeek V4 Flash or V4 Pro as a Claude Code-style agent: 1M context, tool calling, from $0.15 per million input tokens. Setup, model ids and prices.