dani-free

Imported from somdipto/Dani-Free-proxy (skills/dani-free/SKILL.md). Install upstream with npx skills add somdipto/Dani-Free-proxy --skill dani-free. Copyright stays with the author.

name: dani-free description: Operate the Dani-Free local OpenAI-compatible router. auto starts at opencode/nemotron-3-ultra-free and walks the OpenCode-first six-model failover chain. Canonical agent integration guide: the package root AGENT_INTEGRATION.md. Use it for provider setup, protocol behavior, the six-model allowlist, OMP integration, SDK examples, failure handling, and the completion checklist.

Use this skill to configure and operate the Dani-Free package; the package itself owns backend routing and protocol translation. Do not implement a second router in the skill.

Start

1. Install Bun (1.1 or newer) and ensure each selected backend is installed and configured.
2. From a checkout, run ./install.sh, or run bun install and bun run start. No checkout yet: run curl -fsSL https://raw.githubusercontent.com/somdipto/Dani-Free-proxy/main/bootstrap.sh | bash (installs Bun if missing, clones, runs the full setup).
3. Check GET /health and GET /v1/models.
4. Send a minimal non-streaming request to POST /v1/chat/completions before enabling advanced features.

The default listener is 127.0.0.1:4190. Set DANI_FREE_HOST and DANI_FREE_PORT to change it. Keep the listener on loopback unless access from another trusted machine is intentional.

Model selectors and routing

• auto starts at opencode/nemotron-3-ultra-free and walks the OpenCode-first six-model failover chain. Retryable failures (transport errors, 408, 429 with a brief backoff, 5xx, HTTP 200 with empty content, invalid JSON, an {"error": …} envelope — OpenAI-style object, bare-string, or string-array form — or an oversize body) advance to the next model; other 4xx refusals are passed through as-is. A request-deadline timeout or client cancellation never advances: it ends the request with a 504.
• Standard listener ids: three OpenCode free models, then three Kilo free models.
• Other kilo/<id>, mimo/<id>, and opencode/<id> values are rejected unless a custom adapter set was supplied.
• Use an exact model id returned by /v1/models. An explicit selector is tried first, then the remaining models in the chain: explicit selectors fail over too, they do not fail closed.
• Advertised models include tools and reasoning so OMP coding requests do not 422.

The API is OpenAI-compatible at /v1/models and /v1/chat/completions. Client cancellation is propagated and terminal.

Backend configuration

Kilo Code requires its configured gateway and API key when required by the endpoint. OpenCode ids require the local :4187 proxy.

See:

• Routing for selector and request examples.
• Backend configuration for every environment variable and verification procedure.
• Diagnostics for health checks, failure interpretation, and safe troubleshooting.

Security rules

• Never commit .env, API keys, cookies, bearer tokens, or command output containing secrets.
• Keep .env readable only by the local user (chmod 600 .env).
• Prefer loopback host binding. If binding beyond loopback, put the router behind authenticated transport and a firewall; do not treat DANI_FREE_API_KEY as a substitute for network controls.
• Redact Authorization headers and full request/response bodies before sharing diagnostics.
• Treat DANI_FREE_MIMO_COMMAND as code: use an absolute, trusted executable and avoid shell interpolation of untrusted request data.

Current limitation

Free-tier availability can change. If Kilo returns 429/503/timeout, Dani-Free reports that failure and does not switch models.