[LIVE] [CryptoKR123/hydra-free-runner] 3 endpoint(s)
Gemini
API Proxy/Gemini
2,223 characters
Status: [LIVE]
Endpoints found:
https://ai-gateway.kilo.ai/api/v1
https://ai-gateway.kilo.ai/api/v1/models
https://oai.endpoints.kepler.ai.cloud.ovh.net/v1/models
Source: https://github.com/CryptoKR123/hydra-free-runner
# Hydra free runner
A small, standard-library Python runner for public research questions on GitHub
Actions. It contains no wallets, strategy code, private state, or knowledge vault.
It never loads local credential files. Paper only; executor NO_TRADE.
Dispatch `free-lane-scan.yml` on `main` with `workgroups` (base64 JSONL), `ids`
(a JSON array in the same order), `request_id`, and `max_calls` (1–12).
Alternatively provide `question` and space-separated `urls`.
Each workgroup contains `question`, optional `urls`, optional `providers`, and
optional `max_calls`. Nonempty `docs` is refused before any provider call.
Inputs and outputs are public: callers must remove all private material first.
URLs are unverified leads, not fetched sources; answers are never marked grounded.
The restricted cloud free profile uses these repository Actions secret names:
All 15 names are mapped in the workflow: NVIDIA/NVIDIA2, OpenRouter, Gemini,
Groq/Groq2, Mistral/Mistral2, Cerebras, SambaNova, HuggingFace,
Cloudflare/Cloudflare2 plus the Cloudflare account ID, and Jina.
`runner/lanes.json` binds each name to its fixed provider endpoint and adapter.
Jina retrieves the first supplied URL and is counted as source capture, never as
a functional model answer. Cloudflare requires a valid account ID.
OpenRouter uses its free router and a zero-price provider ceiling.
No paid or subscription fallback exists. Configure only free-tier credentials;
this runner does not change account billing settings or replenish quotas.
Heads run concurrently, rotate their starting provider, and serialize each
provider through a shared lock within the batch. Each head visits an account lane at
most once, stops after `max_calls` successful answers, and considers at most
`2 * max_calls` actual calls (missing secrets do not consume this budget). Missing secrets, HTTP failures, unsupported providers,
and deadline exhaustion produce typed records. Requests have a 30-second timeout,
256 KiB response cap, and a 48