Skip to content
National Research Platform NRP Documentation

Client Configurations

Pre-built configurations for popular coding CLIs. Every config here assumes:

  • Endpoint: https://ellm.nrp-nautilus.io/v1, Anthropic-compatible at /anthropic
  • Token: the value from the LLM token page, passed in Authorization: Bearer … or whatever the client’s API-key field expects

API Access covers getting a token, and cache_salt for tenants whose prompts should not be cached across users.

VS Code

https://code.visualstudio.com

VS Code reaches NRP models through Chat: Manage Language Models, which adds them to the Copilot Chat model picker.

  1. Command Palette (Ctrl+Shift+P / Cmd+Shift+P) → Chat: Manage Language Models
  2. Add Models → Custom Endpoint
  3. Endpoint URL: https://ellm.nrp-nautilus.io/v1/chat/completions

The apiKey below uses VS Code’s ${input:...} substitution, so VS Code asks for your token the first time you use a model and stores it securely from then on.

Paste into the Add Models prompt 3 models 
{
  "name": "NRP",
  "vendor": "customendpoint",
  "apiKey": "${input:chat.lm.secret.2630e22e}",
  "apiType": "chat-completions",
  "models": [
    {
      "id": "glm-5",
      "name": "glm-5",
      "url": "https://ellm.nrp-nautilus.io/v1/chat/completions",
      "toolCalling": true,
      "vision": false,
      "maxInputTokens": 1048576,
      "maxOutputTokens": 100000
    },
    {
      "id": "qwen3",
      "name": "qwen3",
      "url": "https://ellm.nrp-nautilus.io/v1/chat/completions",
      "toolCalling": true,
      "vision": true,
      "maxInputTokens": 1000000,
      "maxOutputTokens": 100000
    },
    {
      "id": "kimi",
      "name": "kimi",
      "url": "https://ellm.nrp-nautilus.io/v1/chat/completions",
      "toolCalling": true,
      "vision": true,
      "maxInputTokens": 131072,
      "maxOutputTokens": 100000
    }
  ]
}

omp

https://omp.sh (oh-my-pi), whose custom-model reference covers the fields below.

~/.omp/agent/models.yml 2 models 
providers:
  nrp:
    baseUrl: https://ellm.nrp-nautilus.io/v1
    api: openai-completions
    apiKey: <YOUR_API_KEY>
    models:
      - id: kimi
        name: kimi
        reasoning: true
        supportsTools: true
        input: [text, image]
        contextWindow: 131072
        omitMaxOutputTokens: true
      - id: qwen3-small
        name: qwen3-small
        reasoning: true
        thinking:
          mode: effort
          efforts: [low, medium, xhigh]
          defaultLevel: xhigh
        supportsTools: true
        input: [text, image]
        contextWindow: 1000000
        omitMaxOutputTokens: true

OpenCode

https://github.com/anomalyco/opencode

Ctrl+P → Switch models finds NRP; Switch model variant picks a thinking level. Either set OPENAI_API_KEY or replace {env:OPENAI_API_KEY} with the key itself.

~/.config/opencode/opencode.jsonc qwen3 as the default model

"model" takes a provider/model id, so the default is NRP/qwen3. "small_model" is what OpenCode uses for side tasks such as session titles — qwen3-small keeps those off the largest GPUs. An editor validating against opencode.ai/config.json underlines both values, because the schema autocompletes from the models.dev catalog and cannot know about a custom provider; OpenCode resolves them from the provider block.

{
  "$schema": "https://opencode.ai/config.json",
  "model": "NRP/qwen3",
  "small_model": "NRP/qwen3-small",
  "provider": {
    "NRP": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "NRP",
      "options": {
        "baseURL": "https://ellm.nrp-nautilus.io/v1",
        "apiKey": "{env:OPENAI_API_KEY}"
      },
      "models": {
        "qwen3": {
          "name": "qwen3",
          "limit": {
            "context": 1000000,
            "output": 65536
          },
          "modalities": {
            "input": ["text", "image", "video"],
            "output": ["text"]
          },
          "tool_call": true,
          "reasoning": true,
          "attachment": true
        },
        "qwen3-small": {
          "name": "qwen3-small",
          "limit": {
            "context": 1000000,
            "output": 65536
          },
          "tool_call": true,
          "reasoning": true
        }
      }
    }
  }
}

"limit.output" caps a single response, not the context window. Never set it close to or above "context".

~/.config/opencode/opencode.jsonc every model in the catalog

Nine models, each with its context window, modalities, and thinking controls. qwen3-embedding is an embedding model rather than a chat model, so it is not listed.

Thinking is selected per model with Ctrl+P → Switch model variant, and the key differs by model: qwen3 and qwen3-small take a top-level reasoning_effort (low, medium, or xhigh, the default) and turn thinking off through chat_template_kwargs.enable_thinking; glm-5 and deepseek-v4-flash take reasoning_effort inside chat_template_kwargs (low, high, or max, the default), and only deepseek-v4-flash can turn thinking off, with thinking: false. Staying on a model’s default variant leaves the server-side default in place. gemma-small is served without a reasoning parser, so its thinking arrives inline in the answer rather than as a separate reasoning block.

{
  "$schema": "https://opencode.ai/config.json",
  "model": "NRP/qwen3",
  "small_model": "NRP/qwen3-small",
  "provider": {
    "NRP": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "NRP",
      "options": {
        "baseURL": "https://ellm.nrp-nautilus.io/v1",
        "apiKey": "{env:OPENAI_API_KEY}"
      },
      "models": {
        "qwen3": {
          "name": "qwen3",
          "limit": {
            "context": 1000000,
            "output": 65536
          },
          "modalities": {
            "input": ["text", "image", "video"],
            "output": ["text"]
          },
          "tool_call": true,
          "reasoning": true,
          "attachment": true,
          "variants": {
            "low": {
              "reasoning_effort": "low"
            },
            "medium": {
              "reasoning_effort": "medium"
            },
            "instruct": {
              "chat_template_kwargs": {
                "enable_thinking": false
              }
            }
          }
        },
        "qwen3-small": {
          "name": "qwen3-small",
          "limit": {
            "context": 1000000,
            "output": 65536
          },
          "modalities": {
            "input": ["text", "image", "video"],
            "output": ["text"]
          },
          "tool_call": true,
          "reasoning": true,
          "attachment": true,
          "variants": {
            "low": {
              "reasoning_effort": "low"
            },
            "medium": {
              "reasoning_effort": "medium"
            },
            "instruct": {
              "chat_template_kwargs": {
                "enable_thinking": false
              }
            }
          }
        },
        "glm-5": {
          "name": "glm-5",
          "limit": {
            "context": 1048576,
            "output": 65536
          },
          "modalities": {
            "input": ["text"],
            "output": ["text"]
          },
          "tool_call": true,
          "reasoning": true,
          "variants": {
            "low": {
              "chat_template_kwargs": {
                "reasoning_effort": "low"
              }
            },
            "high": {
              "chat_template_kwargs": {
                "reasoning_effort": "high"
              }
            }
          }
        },
        "deepseek-v4-flash": {
          "name": "deepseek-v4-flash",
          "limit": {
            "context": 1048576,
            "output": 65536
          },
          "modalities": {
            "input": ["text", "image"],
            "output": ["text"]
          },
          "tool_call": true,
          "reasoning": true,
          "attachment": true,
          "variants": {
            "low": {
              "chat_template_kwargs": {
                "reasoning_effort": "low"
              }
            },
            "high": {
              "chat_template_kwargs": {
                "reasoning_effort": "high"
              }
            },
            "instruct": {
              "chat_template_kwargs": {
                "thinking": false
              }
            }
          }
        },
        "kimi": {
          "name": "kimi",
          "limit": {
            "context": 131072,
            "output": 32768
          },
          "modalities": {
            "input": ["text", "image", "video"],
            "output": ["text"]
          },
          "tool_call": true,
          "reasoning": true,
          "attachment": true
        },
        "minimax-m2": {
          "name": "minimax-m2",
          "limit": {
            "context": 204800,
            "output": 32768
          },
          "modalities": {
            "input": ["text"],
            "output": ["text"]
          },
          "tool_call": true,
          "reasoning": true
        },
        "gpt-oss": {
          "name": "gpt-oss",
          "limit": {
            "context": 131072,
            "output": 32768
          },
          "modalities": {
            "input": ["text"],
            "output": ["text"]
          },
          "tool_call": true,
          "reasoning": true
        },
        "gemma": {
          "name": "gemma",
          "limit": {
            "context": 262144,
            "output": 32768
          },
          "modalities": {
            "input": ["text", "image", "video"],
            "output": ["text"]
          },
          "tool_call": true,
          "reasoning": true,
          "attachment": true,
          "variants": {
            "instruct": {
              "chat_template_kwargs": {
                "enable_thinking": false
              }
            }
          }
        },
        "gemma-small": {
          "name": "gemma-small",
          "limit": {
            "context": 262144,
            "output": 32768
          },
          "modalities": {
            "input": ["text", "image", "video"],
            "output": ["text"]
          },
          "tool_call": true,
          "reasoning": true,
          "attachment": true,
          "variants": {
            "instruct": {
              "chat_template_kwargs": {
                "enable_thinking": false
              }
            }
          }
        }
      }
    }
  }
}

To pin options that should apply on every request to a model — cache_salt, preserve_thinking, or a fixed effort — add an "options" block alongside "name". Anything in "options" is sent with each request; anything in "variants" only while that variant is selected:

        "deepseek-v4-flash": {
          "name": "deepseek-v4-flash",
          "options": {
            "chat_template_kwargs": {
              "thinking": true,
              "preserve_thinking": true
            }
          },
          // ...
        },

Crush

https://github.com/charmbracelet/crush

Search for NRP in the model list once the config is in place. See also Andrea Zonca’s writeup, Configure NRP LLM with OpenCode and Crush.

~/.config/crush/crush.json 1 model — repeat the block per model

Adjust "default_max_tokens" as needed; never push it close to or above "context_window". Use "enable_thinking" or "thinking" as the model requires.

{
  "$schema": "https://charm.land/crush.json",
  "options": {
    "disable_metrics": true,
    "disable_provider_auto_update": false,
    "debug": false,
    "debug_lsp": false,
    "attribution": {
      "trailer_style": "none",
      "generated_with": false
    }
  },
  "mcp": {},
  "providers": {
    "nrp": {
      "name": "NRP",
      "type": "openai-compat",
      "base_url": "https://ellm.nrp-nautilus.io/v1",
      "api_key": "<LLM_API_KEY>",
      "models": [
        {
          "id": "kimi",
          "name": "kimi",
          "context_window": 131072,
          "default_max_tokens": 32768,
          "modalities": {
            "input": ["text", "image", "video"],
            "output": ["text"]
          }
        }
      ]
    }
  }
}

For "chat_template_kwargs" or other "extra_body" parameters, add this within each model, at the same level as "models":

      "extra_body": {
        "chat_template_kwargs": {
          "thinking": true,
          "preserve_thinking": true
        }
      },

pi

https://pi.dev

Ctrl-L switches models — search for NRP, or type /model modelname.

~/.pi/agent/models.json 2 models

apiKey accepts the literal key, an environment variable name (OPENAI_API_KEY), or a shell command (!decrypt_key NRP).

{
  "providers": {
    "nrp": {
      "baseUrl": "https://ellm.nrp-nautilus.io/v1",
      "api": "openai-completions",
      "apiKey": "<YOUR_API_KEY>",
      "models": [
        {
          "id": "kimi",
          "name": "kimi",
          "input": ["text", "image"],
          "contextWindow": 131072,
          "reasoning": true
        },
        {
          "id": "qwen3-small",
          "name": "qwen3-small",
          "input": ["text", "image"],
          "contextWindow": 1000000,
          "reasoning": true,
          "compat": {
            "thinkingFormat": "qwen-chat-template",
            "supportsDeveloperRole": false
          }
        }
      ]
    }
  }
}

Kimi CLI

https://github.com/MoonshotAI/kimi-cli

~/.kimi/config.toml 1 model 
default_model = "kimi"
default_thinking = true
default_yolo = false
default_plan_mode = false
show_thinking_stream = true
merge_all_available_skills = true
telemetry = false


[providers.nrp]
type = "openai_legacy"
base_url = "https://ellm.nrp-nautilus.io/v1"
api_key = "<YOUR_API_KEY>"


[models.kimi]
provider = "nrp"
model = "kimi"
max_context_size = 131072
capabilities = ["thinking", "image_in", "video_in"]


[background]
max_running_tasks = 4
keep_alive_on_exit = false
agent_task_timeout_s = 3600

Claude Code

https://github.com/anthropics/claude-code

Claude Code talks to NRP through the Anthropic-compatible endpoint at /anthropic.

~/.claude/settings.json any model, by name 
{
  "env": {
    "ANTHROPIC_BASE_URL": "https://ellm.nrp-nautilus.io/anthropic",
    "ANTHROPIC_AUTH_TOKEN": "<llm-token>",
    "ANTHROPIC_MODEL": "<name-of-nrp-model>",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "<name-of-nrp-model>",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "<name-of-nrp-model>",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "<name-of-nrp-model>",
    "ANTHROPIC_DEFAULT_FABLE_MODEL": "<name-of-nrp-model>",
    "CLAUDE_CODE_SUBAGENT_MODEL": "<name-of-nrp-model>",
    "ENABLE_TOOL_SEARCH": "false",
    "CLAUDE_CODE_AUTO_COMPACT_WINDOW": "<context-size-of-model>",
    "CLAUDE_CODE_EFFORT_LEVEL": "max",
    "CLAUDE_STREAM_IDLE_TIMEOUT_MS": "3000000",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1",
    "CLAUDE_CODE_DISABLE_FEEDBACK_SURVEY": "1",
    "CLAUDE_CODE_ENABLE_TELEMETRY": "0",
    "DISABLE_TELEMETRY": "1",
    "API_TIMEOUT_MS": "3000000",
    "CLAUDE_CODE_MAX_RETRIES": "10"
  }
}

Copilot CLI

https://github.com/github/copilot-cli

Environment variables any model, by name

Set COPILOT_PROVIDER_MAX_PROMPT_TOKENS to the model’s context length and COPILOT_PROVIDER_MAX_OUTPUT_TOKENS to a smaller value — roughly 1/16 to 1/4 of the context, depending on the task.

Terminal window
COPILOT_PROVIDER_BASE_URL="https://ellm.nrp-nautilus.io/v1"
COPILOT_PROVIDER_KEY="<llm-token>"
COPILOT_MODEL="<name-of-nrp-model>"
COPILOT_PROVIDER_MAX_PROMPT_TOKENS="<context-length>"
COPILOT_PROVIDER_MAX_OUTPUT_TOKENS="<max-output>"
U.S. National Science Foundation

This work was supported in part by National Science Foundation (NSF) awards CNS-1730158, ACI-1540112, ACI-1541349, OAC-1826967, OAC-2112167, CNS-2100237, CNS-2120019.