Back to the map
Agents & Multi-agent Systems2026
Claude on NIM
An autonomous coding agent that runs on free GPU endpoints and verifies its own code.
Overview
The Claude Agent SDK speaks the Anthropic Messages API; NVIDIA's free NIM endpoints speak OpenAI. Claude on NIM bridges them with a LiteLLM proxy so you get an interactive Claude CLI (persistent tmux sessions that survive disconnects, mid-conversation model switching, and remote control over Tailscale) running inference on models like GLM, Kimi, and gpt-oss at no cost. Its autonomous mode wires a PostToolUse hook that runs linters, type-checkers, and tests on generated code and feeds failures back into the agent loop, so unattended runs self-correct instead of drifting.
Highlights
- Routes the Agent SDK onto NIM via ANTHROPIC_BASE_URL, no SDK fork, with a validated proxy that translates Anthropic to NIM.
- Deterministic self-verification: hooks always fire, unlike system-prompt requests the model can ignore.
- Containerized autonomy with git credentials and safety guardrails (denylist plus path confinement) for real commits and pushes.