@conclave-ai/agent-ollama

Conclave AI Ollama agent. Routes reviews to a local or self-hosted Ollama instance via its OpenAI-compatible endpoint.

Install

Pulled in automatically by @conclave-ai/cli when "ollama" is in .conclaverc.json → agents.

Standalone:

pnpm add @conclave-ai/agent-ollama

Why Ollama

• Zero API cost. Compute is user-paid hardware; the wire is free.
• Zero API key. Nothing leaves the machine if the Ollama instance is local.
• Any model you've pulled. ollama pull llama3.3, qwen3, deepseek-r1, gpt-oss, etc. — whatever your hardware runs.

Prerequisites

# Install Ollama: https://ollama.com/download
ollama serve                 # starts daemon on localhost:11434
ollama pull llama3.3         # or whichever model the config names

Env / options

| Var / option | Default | Notes | |---|---|---| | OLLAMA_BASE_URL env / baseURL ctor opt | http://localhost:11434/v1 | Override to point at a remote / container / non-default port | | model ctor opt | llama3.3 | Any model ollama list shows |

No API key required. The underlying openai SDK gets the literal string "ollama" as a placeholder so its constructor doesn't throw.

Usage

import { OllamaAgent } from "@conclave-ai/agent-ollama";

const agent = new OllamaAgent({ gate: efficiencyGate });
const result = await agent.review(ctx);
// result.costUsd === 0 always — local inference is not metered

Caveats

• actualCost always returns 0. If you want wall-clock or kWh accounting, layer it on top of MetricsRecorder.latencyMs; that lives outside v0.1 scope.
• deepseek-r1 and other reasoning models emit far more output tokens than the default maxTokens: 8192 estimate. Bump maxTokens via constructor if you route reasoning-heavy reviews through Ollama.
• Structured JSON output depends on the model. Smaller models may fail the strict schema; the CLI surfaces parse errors like any other agent.