Agents

Crush

Project-local Crush v0.100.0 configuration.

Setting up with an AI assistant? Copy the setup prompt or open this guide as Markdown.

Setup prompt

Read https://zrelay.net/docs/ai.md and https://zrelay.net/docs/md/agents/coding-agents/crush.md first.

Configure my installed Charm Crush to use Zrelay in this project. Detect OS, shell and Crush version; ask before installing or upgrading anything. Read official Crush documentation/source for that version and https://keys.zrelay.net/api/catalog. Preserve existing configuration, providers, authentication and permissions; merge only a new uniquely named Zrelay provider. For v0.100.0 use project crushrc builtins, type responses, base https://api.zrelay.net/v1 and model gpt-6.1-sol only if enabled for my GPT key's pool. Set both large and small slots to that provider, context budget 128000 and maximum output 4096 or lower catalog limits. Do not migrate legacy JSON unnecessarily. Read ZRELAY_API_KEY at runtime; provide local hidden-entry instructions appropriate to my shell. Never request, read aloud, print, log or embed the real key in prompts, files or command arguments. Inspect the final diff and perform only offline syntax/configuration checks without expanding secrets. Do not claim connectivity, streaming or tools are verified. Ask for explicit consent before any paid inference; keep any authorized smoke test bounded and test streaming/tools separately. Report exact files changed and untested limitations.

Manual setup

1. Check crush --version. For v0.100.0 merge the first block into project .crushrc, or preserve an existing crushrc. These are configuration-file builtins, not terminal commands. The zrelay-gpt provider ID must be unused; the final two lines select the project large and small slots.

2. Enter ZRELAY_API_KEY locally and launch crush from that environment. The embedded Bash configuration works on Windows too; the PowerShell example requires PowerShell 7. Global alternative: ~/.config/crush/crushrc, respecting XDG and overrides.

3. Use a GPT Flex/Prime key for gpt-6.1-sol Responses. Retain old JSON configuration if present rather than migrating automatically. Current provider types are responses, completions and messages.

4. Claude variant: add zrelay-claude with type messages, base https://api.zrelay.net, and claude-sonnet-5-5 with a matching Claude key. This adapter uses the Anthropic SDK, which adds v1/messages; do not copy the /v1 base from OpenCode or Continue.

Configuration and commands

bash

provider add zrelay-gpt --name Zrelay --type responses \
  --base-url "https://api.zrelay.net/v1" \
  --api-key "${ZRELAY_API_KEY:?set ZRELAY_API_KEY}"
model add zrelay-gpt/gpt-6.1-sol --name "GPT 6.1 Sol" \
  --context-window 128000 --default-max-tokens 4096
model large zrelay-gpt/gpt-6.1-sol --max-tokens 4096
model small zrelay-gpt/gpt-6.1-sol --max-tokens 4096

bash

read -rsp 'Zrelay API key: ' ZRELAY_API_KEY; printf '\n'
export ZRELAY_API_KEY
crush

powershell

$env:ZRELAY_API_KEY = Read-Host 'Zrelay API key' -MaskInput
crush

Alternative: ask an AI assistant

Copy this complete prompt into your coding assistant. It describes configuration work; installation and paid tests need your authorization.

Setup prompt

Configure my installed Charm Crush to use Zrelay in this project. Detect OS, shell and Crush version; ask before installing or upgrading anything. Read official Crush documentation/source for that version and https://keys.zrelay.net/api/catalog. Preserve existing configuration, providers, authentication and permissions; merge only a new uniquely named Zrelay provider. For v0.100.0 use project crushrc builtins, type responses, base https://api.zrelay.net/v1 and model gpt-6.1-sol only if enabled for my GPT key's pool. Set both large and small slots to that provider, context budget 128000 and maximum output 4096 or lower catalog limits. Do not migrate legacy JSON unnecessarily. Read ZRELAY_API_KEY at runtime; provide local hidden-entry instructions appropriate to my shell. Never request, read aloud, print, log or embed the real key in prompts, files or command arguments. Inspect the final diff and perform only offline syntax/configuration checks without expanding secrets. Do not claim connectivity, streaming or tools are verified. Ask for explicit consent before any paid inference; keep any authorized smoke test bounded and test streaming/tools separately. Report exact files changed and untested limitations.

Limits and verification

Keep both slots at 4096 output and a conservative context budget of 128000. The configuration interpolates the environment at runtime: do not expand secrets into a saved file. Complete tool and streaming behavior is untested.

These recipes were reviewed against documentation and public catalog data. No client installation, authenticated request, paid inference, streaming or tool loop was executed. Validate configuration first; a later short text test does not establish complete agent compatibility.

Recheck the live catalog and your key’s provider/pool. The reviewed examples have 128000 input and 4096 output maxima unless a client limitation above prevents controlling the request. Never print, log or commit credentials.