Priority: P2 Status: ✅ done Assignee: chris Created: 2026-04-29 Completed: 2026-05-10

Context#

GitHub Models (via models.inference.ai.azure.com) provides a free tier for multiple models including GPT-4o-mini, GPT-4o, DeepSeek-R1, Phi-4, Llama 3.1/3.3, Cohere Command R+, and others. Access is via a GitHub Personal Access Token (PAT) with no special scopes needed — the token just needs to exist.

Acceptance Criteria#

• 1. Provider registered in openclaw.json
• 2. Smoke test: CLI inference works with github-models provider
• 3. Models listed in MC Models page

Implementation#

1. Provider config in openclaw.json#

Added github-models provider with 9 models:

• gpt-4o-mini (128k ctx, free, 1440 req/day)
• gpt-4o (128k ctx, free, 200 req/day)
• DeepSeek-R1 (64k ctx, free, reasoning, 200 req/day)
• Phi-4 (128k ctx, free, 500 req/day)
• Meta-Llama-3.1-405B-Instruct (131k ctx, free, 200 req/day)
• Meta-Llama-3.1-8B-Instruct (131k ctx, free, 1440 req/day)
• Llama-3.3-70B-Instruct (131k ctx, free, 200 req/day)
• Cohere-command-r-plus-08-2024 (128k ctx, free, 200 req/day)
• ministral-3b (32k ctx, free, 1440 req/day)

Config block added to ~/.openclaw/openclaw.json:

"github-models": {
  "baseUrl": "https://models.inference.ai.azure.com",
  "apiKey": { "source": "env", "provider": "default", "id": "GITHUB_TOKEN" },
  "api": "openai-completions",
  "models": [ ... ]
}

2. Smoke test results#

$ openclaw capability model run --model "github-models/gpt-4o-mini" --prompt "Say hi"
→ "Hello!" ✅

$ openclaw capability model run --model "github-models/DeepSeek-R1" --prompt "2+2?"  
→ "4" ✅

3. MC Models page#

Provider health shows "github-models": "ok" — all 9 models visible in the models endpoint.

Auth#

The GITHUB_TOKEN env var is stored in ~/.openclaw/.env and loaded by the mission-control systemd service via EnvironmentFile=-/home/maxime/.openclaw/.env. The token is a 40-char ghp_* token and is confirmed working (tested with live inference).

Blocked on (resolved)#

Maxime needs to verify/update his GitHub PAT to include models:read scope.

• Resolution: The GITHUB_TOKEN was already set in ~/.openclaw/.env but not sourced in the agent runtime. The token is valid and working. The service already loads this via systemd EnvironmentFile.
• Importantly: This agent (Chris) doesn't automatically source ~/.openclaw/.env — if you want Chris to use github-models as his default model, the env needs to be sourced at agent launch or the openclaw config needs to use a source: env with id: GITHUB_TOKEN (which it does).

🕵️ Review — Debbie verified 2026-05-10T23:05Z.

Maxime flagged: "Chris shouldn't change to another model. This task is to setup the models so we can use them."

Findings: No agent has its primary model set to github-models. The provider is purely available as a fallback option (referenced in agents.defaults.model.fallbacks as one of several fallbacks after deepseek-reasoner and gemini-2.5-flash). Chris's agent config shows model: deepseek-chat — unchanged. No model assignments were modified. The provider is ready for use when anyone chooses to reference it, but no automatic switching has occurred.

Resolution: ✅ No changes needed. The task was correctly implemented by Chris.

📌 Re-review by debbie (2026-05-10): Confirmed the provider is live, configured, and working. Smoke-tested both gpt-4o-mini and DeepSeek-R1 successfully. Run log: .runs/r-t-3339-1746918301136.log.