ollama-agent

An extensible agent platform in Go that works exclusively with Ollama and is optimized for small local LLMs. It distributes requests across multiple Ollama servers, picks the right model for each task on its own, and ships with a dark web UI.

Features

Operation across multiple servers

• Endpoint pool: use multiple Ollama servers in parallel, with health checks, automatic failover, and recovery. Requests only go to servers that actually have the required model.
• Docker: a container image that starts the web UI directly; configuration optionally entirely via environment variables.

The right model for each task

• Automatic model selection: task categories (chat, code, summarize, …) are mapped to model preference lists via JSON config; a small router model classifies tasks when no category is specified.
• Built-in benchmarks: a standardized test measures capability tiers, real context size, speed, and stability per model and server — the context depth test additionally checks whether a model truly understands its large context.

Robust with small models

• Agent loop with tool calling: built against the typical weaknesses of small models — salvage parsing of malformed tool calls, short prompts, tool subsetting by category, loop detection, context compaction.
• Patient timeouts: local models are allowed to take minutes for a response without anything getting aborted.

Interface

• Web UI: multiple chat sessions with automatic naming, full-text search, per-response token display, context meter with a compress function, job overview, model matrix with benchmark launch, and a settings page.
• Read-aloud & voice input: optional via Google Cloud — speaker icon under each response, microphone on the input field.
• CLI: the same platform, interactive or scriptable, on the command line.

Extensible

• Markdown skills: capabilities as SKILL.md folders, extensible without recompiling.
• MCP client: connect external MCP servers (stdio + HTTP), their tools become available to the agent.
• Web research: built-in web search and page analysis with a compression layer, so even small models can work with web content.
• Cron scheduler: recurring agent jobs via cron expression.

Quick start

make build
./bin/ollama-agent config init > config.json   # generate example config, then adjust
./bin/ollama-agent ui                          # start the daemon and open the web UI

Full details in Installation. If you’d rather work on the command line, get started at Command Line.

Where to go next?

I want to …	Page
install the platform	Installation
understand config.json	Configuration
get to know the interface	Web UI
look up all commands	Command Reference
pick the right model	Model Guide
add custom capabilities	Extending the Platform
contribute to the core	Architecture
look up every endpoint with schema and example	API Reference
know what changed recently	Changelog