Coding Agents
Pawan.Krd speaks three wire formats — OpenAI Chat Completions, OpenAI Responses and Anthropic Messages — with streaming tool calls, parallel tool calls, structured outputs and reasoning. That covers the common coding agents. Pick a model with the toolCalling capability (shown on the Models page).
Claude Code
Claude Code talks to /v1/messages. Point it at Pawan.Krd and choose which models it uses for the main loop and for quick background tasks.
Shell
export ANTHROPIC_BASE_URL="https://api.pawan.krd"
export ANTHROPIC_AUTH_TOKEN="pk-your_key"
export ANTHROPIC_MODEL="openai/gpt-oss-120b"
export ANTHROPIC_SMALL_FAST_MODEL="openai/gpt-oss-20b"
claudeCodex CLI
Codex can use either the Responses API or Chat Completions.
config.toml
# ~/.codex/config.toml
model = "openai/gpt-oss-120b"
model_provider = "pawan"
[model_providers.pawan]
name = "Pawan.Krd"
base_url = "https://api.pawan.krd/v1"
env_key = "PAWAN_API_KEY"
wire_api = "responses" # or "chat"Cline, Roo Code, Continue, Aider, OpenCode, Cursor and others
Choose the "OpenAI compatible" provider and use:
Settings
Base URL: https://api.pawan.krd/v1
API key: pk-your_key
Model: any chat model ID with tool calling, e.g. openai/gpt-oss-120bWhat to expect
- Long conversations are trimmed to the model's context window; a tool call and its result are always kept or dropped together.
- Token counting endpoints (
/v1/messages/count_tokens,/v1/responses/input_tokens) are free and don't count toward the per-minute rate limit. - If a stream is interrupted after output was produced, you are billed for what was generated; if nothing was produced, the hold is refunded.
- Prompt size is billed from the provider-reported token count, which includes tool definitions.
Hosted tools
Provider-hosted tools (web search, code execution, computer use, Anthropic server tools) are not available. Agents that define their own tools — which is how all of the clients above work — are fully supported.