Feature: Support local OpenAI-compatible providers (Ollama, LM Studio) via localProvider in AgentDefinition

[vraj00222]

Problem

Right now, every LLM call Codebuff makes goes through its backend → OpenRouter, regardless of what model string is set in an AgentDefinition. There is no way to point an agent at a locally-running model.

This blocks three real use cases:

• Privacy — code never leaves the machine
• Offline / air-gapped — no internet, no Codebuff
• Cost splitting — cheap tasks (file picking) locally, complex reasoning on cloud

Related: #288 requested LiteLLM integration. This is a narrower, zero-dependency alternative — no new packages, no Python proxy.

---

Proposed Solution

Add an optional localProvider field to AgentDefinition. When present, the CLI calls the local endpoint directly, bypassing the Codebuff backend entirely.

Type change

// .agents/types/agent-definition.ts

export interface AgentDefinition {
  // ...existing fields unchanged

  localProvider?: {
    baseUrl: string  // e.g. "http://localhost:11434/v1"
    apiKey?: string  // optional — Ollama ignores it
  }
}

Dispatch logic

// cli/src/<agent-dispatch>

if (agent.localProvider) {
  const client = new OpenAI({
    baseURL: agent.localProvider.baseUrl,
    apiKey: agent.localProvider.apiKey ?? 'ollama',
  })
  return client.chat.completions.create({ model: agent.model, messages, stream: true })
}

// Default: existing Codebuff backend → OpenRouter path

openai is already in the dependency tree via OpenRouter — zero new packages needed.

Global fallback (env var)

CODEBUFF_LOCAL_BASE_URL=http://localhost:11434/v1

If set, applies as a fallback localProvider for any agent that doesn't define one explicitly.

---

Scope

• Files touched: ~3 (agent-definition.ts, CLI dispatch, env schema)
• Lines of code: ~80
• New dependencies: None
• Breaking changes: None — fully backward-compatible

---

Happy to Submit a PR for this if the approach looks good.