Skip to content

Developer Tools||2 min read

Running gh-aw agents on NVIDIA free NIM endpoints

Wiring GitHub Agentic Workflows to use NVIDIA's free hosted NIM endpoints via Copilot BYOK engine mode

By Sameer Khan

I configured GitHub's Agentic Workflows framework (gh-aw) to work with NVIDIA's free hosted NIM endpoints. This post documents the exact configuration and the three failures I hit before getting a successful run.

The task: wire the repo's GitHub Agentic Workflows (github/gh-aw) to NVIDIA's free hosted NIM endpoints (integrate.api.nvidia.com, specifically nvidia/nemotron-3.5-lightning-30b-a3b at 40 RPM free tier) via the Copilot BYOK (Bring Your Own Key) engine mode.

Three failures occurred in sequence before a green run:

  1. AWF proxy rejected every request with 400 'Model has no AI credits pricing' - This was fixed by adding frontmatter configuration: models.default-ai-credits-pricing input/output 0.000001 (note that the compiler rejects literal 0, so a tiny value is required)

  2. Proxy token steering silently rewrote the model to claude-sonnet-5 - Resolved by setting sandbox.agent.token-steering: false and model-fallback: false in the configuration

  3. The agentics sample's timeout-minutes: 15 killed a successful 16.3-minute research session - Increased timeout to 30 minutes to accommodate longer research sessions

Additional verified facts:

  • NVIDIA NIM serves the /v1/chat/completions endpoint but returns 404 for /v1/responses and /v1/messages (verified via curl testing)
  • Therefore, the Copilot engine's default completions wire API is the only viable fit
  • The threat-detection step requires the larger free model nvidia/nemotron-3-super-120b-a12b to emit the strict THREAT_DETECTION_RESULT JSON format
  • Sustained sessions can encounter free-tier 403 throttling, which is solved by routing through OpenRouter with NVIDIA BYOK integration (uses NVIDIA quota first, then automatically falls back)
  • The primary diagnosis tool is gh aw audit (run-id) for inspecting workflow executions

Sources consulted:

In short: set the tiny credits-pricing value to get past the AWF proxy, disable token steering and model fallback, and raise the session timeout. That's enough to run gh-aw agents on NVIDIA's free models at no cost — including multi-step reasoning tasks and steps like threat detection that require strict JSON output.