I wanted some of my Buzz agents to do jobs that I do not want to spend Claude tokens on, such as tidying text or explaining what something is. That meant finding a runtime that costs nothing to use. I chose Goose, which is open source, and Ollama, which runs AI models on my own computer.
It took far longer than I expected. This article covers the set-up that worked and every problem I hit on the way, because the problems are where the useful lessons are. It follows on from my first Buzz article and my article about my five Buzz agents.
- A runtime is the program that runs an agent. Goose, Claude Code, Codex and Buzz Agent are the four Buzz showed me as ready.
- Goose and Ollama cost nothing and need no account, but the local model is much weaker than Claude.
- Ollama's default context window is too small for an agent with a long list of tools, so raise it.
- Buzz can show an agent as Goose while it quietly runs on Claude. Check the log to be sure.
- Without a web search tool, a local agent answers from memory only.
What is a runtime in Buzz?
When you give a Buzz agent a job, something has to do the thinking and use the tools. That something is the runtime. In the Buzz settings, under Agents, there is a section called Agent runtimes. On my computer it lists Buzz Agent, Goose, Claude Code and Codex, and all four showed Ready. An Add runtimes button sits underneath for more.
Each agent you make picks one runtime. That choice decides which program runs the agent, and it also decides which paid plan, if any, gets used. My Claude Code agents use my Claude plan. My Codex agents use my ChatGPT plan. A Goose agent can use neither, if you point it at a model on your own computer.
Runtimes I looked at
Buzz talks to runtimes through the Agent Client Protocol, or ACP, so any program that speaks ACP can be added. One third-party write-up of Buzz describes three groups. Goose, Claude Code, Codex and Buzz's own agent are built in. A second group of presets includes Cursor, OpenCode, Amp, Kimi Code and Grok Build. The third group is any ACP program you register yourself. The official ACP registry lists about 43 agents.
I only tested Goose. The rest of this table is what I could find out, so treat it as a starting point and check each project yourself.
| Runtime | Free to install | In Buzz | What I found |
|---|---|---|---|
| Goose | Yes, open source | Built in | Works with local models through Ollama. I tested it. |
| OpenCode | Yes, open source | Preset | Not tested |
| Cursor | No, not open source | Preset | Not tested. I could not confirm what its free plan includes. |
| Gemini CLI | Not confirmed | Not a preset | Google says it was replaced by Antigravity CLI on June 18, 2026 |
| GitHub Copilot CLI | Needs a Copilot plan | Not a preset | Listed in the ACP registry |
A runtime can cost nothing to install and still send your work to a paid model. The runtime is the program. The model is the thinking, and the model is usually where the cost sits.
Why Goose and Ollama
Goose is an open source agent that can read files, run commands and use tools. Its documentation says it has moved to the Agentic AI Foundation. It is general purpose, so it suits jobs like research and writing as well as code. Buzz treats it as a first-class runtime.
Goose needs a model to do the thinking. Ollama is a free program that downloads models to your computer and runs them there, with no account and no data sent anywhere. Together they give me an agent that costs nothing per message.
My computer has 32 GB of memory, a Ryzen 5 5600X processor and an NVIDIA RTX 3050 graphics card with 8 GB. That is enough for models of around 8 billion parameters, which is the small end. Bigger ones would be slow on this machine.
Strong at choosing tools and following long instructions. Uses my Claude plan on every message.
Costs nothing to run and keeps my text on my computer. Noticeably weaker, and slower on my graphics card.
The set-up steps
This is the order that worked in the end, with the problems left out. The problems come next.
Set up Goose with Ollama in seven steps
- Check your computerOpen Task Manager, then the Performance tab, and note your memory and graphics card memory.
- Install OllamaRun the winget command from the first box below in PowerShell.
- Download a modelRun the ollama pull command. Mine is 4.9 GB, so expect a wait.
- Raise the context windowAdd a user environment variable called OLLAMA_CONTEXT_LENGTH, then restart Ollama.
Why raise it?
Ollama starts with a window of 4,096 tokens. Goose's own documentation warns that this is too small and can make tools and instructions get ignored. I used 25,000.
- Tell Goose about OllamaCreate Goose's settings file with the second box below.
- Create the agent in BuzzOn the Agents page choose the plus card. Under AI configuration choose Customize for this agent, set the harness to Goose, the provider to a custom provider with the ID ollama, and the model to a custom model.
- Check the right program is runningLook in the Buzz logs, as shown in the third box below.
winget install --id Ollama.Ollama --source winget ollama pull llama3.1:8b
GOOSE_PROVIDER: ollama GOOSE_MODEL: llama3.1:8b OLLAMA_HOST: localhost:11434
Goose did not create that file for me, so I made it at %APPDATA%\Block\goose\config\config.yaml. I am not sure Buzz reads it. The Goose command line does, and I used the command line for my first tests.
Folder: %APPDATA%\xyz.block.buzz.app\agents\logs Line: buzz-acp starting ... agent_cmd= Goose running: agent_cmd=C:\Users\you\.local\bin\goose.exe acp Claude running: agent_cmd=C:\Users\you\AppData\Roaming\npm\claude-agent-acp.cmd
Running ollama ps in PowerShell also shows which model is loaded and how big its context window is.
What went wrong
Everything above sounds tidy. This is the order I actually found things out in.
1. The small model could not pick the right tool
I added my CPS tool to Goose as a test. The model I started with, qwen3:8b, had to choose from about 90 tools and could not find the one I wanted. It only worked once I raised the context window and named the tool in my request. A small model copes badly with a long list, which is one reason I later took CPS out of Goose again.
2. The agent said Goose and ran Claude
I set the harness to Goose and asked a question. The answer looked fine, but when I read the Buzz log the agent had started Claude Code through its adapter. It was using my Claude plan the whole time. The harness setting had not taken effect, and nothing on the screen told me. The log line in the box above is how I found out.
3. Ollama was not in the provider list
The provider dropdown offered Anthropic, OpenAI, OpenRouter and others, but not Ollama. I chose Custom provider and typed the ID ollama. Buzz saved after every letter I typed, so the Goose log filled with warnings about providers called o, ol, oll and olla. They are harmless leftovers from typing.
4. The model would not load, and my settings kept vanishing
Under Model, Buzz said it could not load live models for ollama and offered built-in options instead. Goose then failed to start with an error saying the setting GOOSE_MODEL could not be found. I added environment variables in the Advanced section of Edit agent, but they disappeared every time I stopped and restarted Buzz. What did save was the Custom model option, where I typed llama3.1:8b. That is what finally worked.
5. The first model started an answer and stopped
I watched qwen3:8b begin to answer and then stop. Later questions seemed to disappear, and Buzz was set to queue messages behind the one still running. Qwen3 is a model that writes out its own reasoning before the real answer, and it may be what tripped things up. I switched to llama3.1:8b, which does not do that. I cannot prove that was the cause.
6. The text I wanted shortened was far too long
I gave the agent a video transcript of about 129 KB, which is many times more than a 25,000 token window holds. A bigger window would use more of my graphics card memory, and my card would start handing work to the processor, which is slower. Short text should work but so far has not and this is still being fixed.
7. It cannot look anything up
I wanted a What's This? agent that searches the internet. In Goose version 1.56 the built-in Computer Controller add-on only handled Word, PDF and Excel files, so there was no search tool. A search add-on means running someone else's code on my computer. I held off, so for now the agent answers from what the model learned in training and says so.
A model answering from memory sounds just as sure when it is wrong. Tell the agent in its instructions to say when it could not check something, and double check anything recent.
When something does not answer, read the logs before changing settings. The Buzz agent log and Goose's own logs told me more in a minute than an hour of guessing.
Where it stands now
Ollama and Goose are installed and work from the command line. The Buzz agent now uses llama3.1:8b, and Ollama loaded that model when I asked my question, so the request does reach my own computer but so far I have not been able to get an answer back from it.
My honest view is that a local 8 billion parameter model is a useful extra and not a replacement. It is free and private, which suits short text jobs, and it is weaker and slower than Claude. If I try again I will start with short text, test one small request at a time and check the log after every change.
Before you trust a new runtime
- Read the Buzz logConfirm the program in agent_cmd is the one you chose.
- Run ollama psConfirm the model and the context size.
- Start with a short questionTry what is an API before any real job.
- Send one message and waitMore messages can queue behind a stuck one.
- Check your settings after a restartMine disappeared once.
FAQ
Do I need an account for Ollama or Goose?
No. Both are free, and neither asked me to sign in. Goose asks you to choose a model provider, and for a local setup that provider is Ollama on your own computer.
Is it really free?
The software is free and each message costs nothing. The cost is your computer's power, the disk space for the model, and some patience, because it is slower than a paid service.
Which model should I download?
Pick one that supports tool use and fits your graphics card memory. I tried qwen3:8b first and settled on llama3.1:8b. Check that your Ollama version meets the model's requirements before downloading.
Can a Goose agent use my content calendar?
It can be added, but I took it out again. A small model struggled to pick the right tool from so many, and the Claude and Codex agents do that job better.
Is a free runtime right for you?
- Which jobs do you give to AI that do not need the strongest model?
- What does your computer have in memory and graphics card?
- Which text would you rather keep on your own computer?
- A runtime runs an agent, and the model behind it decides the real cost.
- Goose and Ollama are free and need no account, but a small local model is weaker than Claude.
- Raise Ollama's context window before you give an agent a long list of tools.
- Check the Buzz log, because an agent set to Goose can still be running Claude.
- Without a search tool, a local agent answers from memory and should say so.
