SambaNova Plugin for Claude Code

Claude Code skills for managing SambaNova models and delegating coding tasks to a sub-agent running on SambaNova Cloud.

Installation

Step 1: Add the marketplace

In Claude Code, run:

/plugin marketplace add <this repository link>

Step 2: Install the plugin

/plugin install sambanova-plugin-cc

Prerequisites

• Python 3.11+ with venv support.
• A SAMBANOVA_API_KEY environment variable — required by /model-info and /code.
• The opencode CLI installed and on your PATH (used by /code; no model configuration of your own is required).

There is no manual setup step. On each session start the plugin builds an isolated virtual environment (.env/) and installs its agent_shims package into it automatically, so the skills are ready to use.

Configuring the base URL

By default the plugin talks to public SambaNova Cloud (https://api.sambanova.ai/v1). If you run a SambaManaged / SambaStack deployment — or any OpenAI-compatible SambaNova endpoint on a different host — point the plugin at it with the SAMBANOVA_BASE_URL environment variable:

export SAMBANOVA_BASE_URL=https://api.example.ai/v1

Notes:

• Set the base URL (ending in /v1). You can also paste a full …/v1/chat/completions endpoint — the trailing /chat/completions is stripped automatically.
• This affects every skill that contacts SambaNova (/code and /model-info).
• The legacy variable SAMBANOVA_API_OVERRIDE is still honored for backward compatibility; SAMBANOVA_BASE_URL takes precedence if both are set.

How it works

The plugin ships an MCP server (mcp_server/server.py, launched via mcp_server/bootstrap.sh and registered in .mcp.json) that exposes each skill as an MCP tool backed by the shared agent_shims Python package.

It maintains its own local SQLite database of model parameters (agent_shims/model_parameters/parameters.db) as the sole source of truth — the plugin never reads your personal opencode config. When you run /code, it generates an isolated, runtime-only opencode configuration for the sub-agent, keeping it cleanly separated from your own setup.

Skills Overview

Skill	Command	Description
code	/code <model> <cwd> <prompt> [--max-tokens <n>] [--session <id>]	Delegate a coding task to a sub-agent
list-models	/list-models	List all models in the local parameters database
model-info	/model-info	Show all models available on the SambaNova platform
update-model	/update-model <name> <ctx> <max_tokens> [params_json]	Add or update a model in the database
reset-model-db	/reset-model-db	Clear all entries from the model database

Skill Details

code

Runs opencode as a sub-agent with a specified model and prompt — useful for delegating tasks like code review, implementation, ideation, or running commands and summarizing the result.

/code <model> <cwd> <prompt> [--max-tokens <n>] [--session <id>]

Arguments:

Argument	Required	Description
model	Yes	Bare model ID from the database (e.g. MiniMax-M2.7, not sambanova/MiniMax-M2.7). Use /list-models to see options.
cwd	Yes	Working directory for the sub-agent. Defaults to $CLAUDE_PROJECT_DIR when omitted; an absolute path is recommended.
prompt	Yes	The prompt to send, quoted as a single argument.
--max-tokens	No	Override the model's max_completion_tokens for this run.
--session	No	Resume a previous /code session (from a prior call) to preserve context. Resuming must use the same cwd as the original call.

The model must already be in the local database (/list-models to check, /update-model to add one). The sub-agent is sandboxed to cwd; grant access to directories outside it only when needed.

list-models

Lists all models currently stored in the local parameters database along with their context length, max completion tokens, and sampling parameters.

/list-models

model-info

Queries the SambaNova API (<base_url>/models, where <base_url> defaults to https://api.sambanova.ai/v1 and is configurable via SAMBANOVA_BASE_URL) to display the full catalog of available models with their context length and max completion tokens. Requires SAMBANOVA_API_KEY to be set.

This shows what models can be used, as opposed to /list-models which shows what is stored locally.

/model-info

update-model

Inserts or updates a model entry in the local parameters database. Look up model details via /model-info and sampling parameters from the model's documentation before writing.

/update-model <name> <context_length> <max_completion_tokens> [sampling_parameters_json]

The sampling parameters argument is a JSON string (e.g. '{"temperature": 0.7, "top_p": 0.9}').

reset-model-db

Deletes all entries from the model parameters database.

/reset-model-db

Architecture

plugins/samba-plugin/
├── .mcp.json                   # registers the plugin's MCP server
├── mcp_server/
│   ├── bootstrap.sh            # ensures the venv, then launches the server
│   └── server.py               # FastMCP server exposing the skills as MCP tools
├── hooks/                      # SessionStart hook warms the venv (no manual setup)
├── skills/                     # one SKILL.md per skill (thin MCP front-ends)
│   ├── code/
│   ├── list-models/
│   ├── model-info/
│   ├── update-model/
│   └── reset-model-db/
└── agent_shims/                # shared Python package (installed into .env)
    ├── model.py                # Model dataclass (id, context_length, max_completion_tokens, sampling_parameters)
    ├── model_parameters/       # SQLite-backed model parameter storage (parameters.db)
    └── opencode/               # opencode runner + injected rules/

Each skill is a thin SKILL.md front-end: it adds the prompting discipline, while the implementation lives in the MCP server's tools, backed by agent_shims.