DocsCoding agents
Coding agents
Run terminal coding agents on your own GPU. Which API each agent speaks, where its configuration lives and which model to use.
Coding agents read your files, run commands and edit code in a loop, sending many requests with tool calls. On Osarai they all run on your own instance, so a long session costs the same hourly price as a short one.
Before you start
- A Ready instance running a model with tool calls, for example
qwen3.8-27b(Quickstart). - An API key in your shell:
Terminal
export OSARAI_API_KEY="qck_your_key_here"PowerShell
$env:OSARAI_API_KEY = "qck_your_key_here"To keep it in new terminals, add the line to ~/.zshrc or ~/.bashrc (on Windows: setx OSARAI_API_KEY "qck_your_key_here").
Compatibility
| Agent | API it uses | Configuration | Tested with |
|---|---|---|---|
| opencode | Chat Completions | opencode.json | opencode 1.19 |
| Claude Code | Anthropic Messages | Environment or ~/.claude/settings.json | Claude Code 2.1 |
| Codex CLI | Responses | ~/.codex/config.toml | Codex CLI 0.167 |
| Crush | Chat Completions | crush.json or crushrc | Crush 0.100 |
| Qwen Code | Chat Completions | Environment or ~/.qwen/.env | Qwen Code 0.27 |
| Aider | Chat Completions | Environment or .aider.conf.yml | Aider 0.86 |
| Goose | Chat Completions | custom_providers/osarai.json | From goose's docs, not run by us |
| Kilo Code CLI | Chat Completions | ~/.config/kilo/kilo.json | Kilo CLI 7.8 |
| Continue CLI | Chat Completions | ~/.continue/config.yaml | Continue CLI 1.5 |
Editors work the same way: see Cline, Roo Code, Continue and Zed.
What every setup has in common
| Setting | Value |
|---|---|
| Base URL (OpenAI format) | https://osarai.xyz/v1 |
| Base URL (Anthropic format) | https://osarai.xyz |
| API key | your qck_… key |
| Model | the id on your instance page, for example qwen3.8-27b |
| Context window | 131072 for qwen3.8-27b on an RTX 5090 (all models) |
Tips for long sessions
- Tell the agent the real context window (each guide shows where), so it compacts before the model runs out of room.
- Lower reasoning effort (
low) for quick edits and raise it for planning; see reasoning effort. - Agents that run commands ask before doing so; review what they want to run.