Harness overview
Harnesses are the reason ctrlyoke can keep one workflow model while talking to very different AI assistants.
See the generated harness compatibility matrix for capability fidelity, evidence, version scope, caveats, and remediation guidance.
At a glance
| Harness | Drives | Notes |
|---|---|---|
copilot-cli | GitHub Copilot CLI (copilot) | Fine-grained tool permission patterns |
claude-cli | Claude Code (claude) | Fine-grained tool permission patterns |
codex-cli | OpenAI Codex CLI (codex) | Sandbox/approval modes, not tool patterns |
opencode-cli | OpenCode (opencode) | Provider-flexible provider/model IDs |
qwen-cli | Qwen Code (qwen) | Fine-grained tool permission patterns |
antigravity-cli | Google Antigravity CLI (agy) | Coarse permission modes only |
Choosing a harness
There is no preferred harness. Use the one you already have installed, subscribe to, or trust with your code — ctrlyoke gives each of them the same prompt syntax, workflow format, output view, and controls. Out of the box the default is Auto-detect: the first installed harness in alphabetical order. Choose a specific one with the Default harness setting in ctrlyoke Settings, override it per workflow, or pick a different harness for each step in the same run.
All harnesses use the same workflow schema and share the core runtime: interactive terminal mode, headless execution, per-harness additionalArgs, CLI session IDs, MCP server attachment, and portable agents and skills. What differs is enforcement of tool permission lists, which hook events fire, per-server MCP tool filtering, and image attachments — check the compatibility matrix before relying on one of those.
See the per-harness pages for setup notes and tradeoffs.
How ctrlyoke talks to a harness
ctrlyoke doesn't bundle, replace, or impersonate any harness. You install each CLI yourself and sign in with that CLI's own login flow, and ctrlyoke calls it for you:
- It runs the vendor's own executable. ctrlyoke starts the CLI you installed (
claude,codex,copilot, and so on) with the arguments that harness expects. The CLI does the model calls, under your account, the same way it would if you typed the command yourself. - It never signs in for you. ctrlyoke has no login for any AI provider, and it never reads or stores a harness's sign-in tokens. You sign in once, in that CLI's own terminal, with its own login flow.
- It never calls a model provider directly. Your prompts and code travel from the harness to its provider, not through ctrlyoke. See Privacy and telemetry.
What ctrlyoke takes off your hands is everything around the call: each harness's flags, session resume, permission modes, MCP server and hook configuration, and its output format. Your use of each harness stays subject to that provider's own terms.