Use Ollama to Power Goose

Run models with Ollama locally and point Goose at Ollama's local provider. Keep coding prompts on-device and avoid API spend.

Index#

Overview
Setup
- Start Ollama locally
- Configure Goose
Beginner usage
Pro usage
Cost savings guide
Privacy guide
Security guide
Appendix

Overview#

Goose supports an Ollama provider for local models running on http://localhost:11434. That lets you keep the same Goose workflow, approvals, extensions, and repo context while running a local model instead of paying for a cloud API on every prompt.

Why Ollama for Goose?

Simple local setup with one background server
No per-token API cost after the model is downloaded
Good fit for private repos or offline-ish workflows
Easy model switching without rebuilding your workflow

Setup#

1) Start Ollama locally#

ollama --version
ollama serve
ollama pull qwen3.5:4b

Quick checks:

curl http://localhost:11434/api/version
ollama ls

If ollama serve is already managed by the desktop app or a service, you only need to make sure the server is listening on localhost:11434.

2) Configure Goose#

Option A: Goose Desktop

Open Goose, go to provider setup, choose Ollama, leave the host as http://localhost:11434, and select an installed local model such as qwen3.5:4b.

Option B: Goose CLI

Run the interactive setup once:

goose configure

Or launch a session with explicit local-provider environment variables:

export GOOSE_PROVIDER=ollama
export OLLAMA_HOST=http://localhost:11434
export GOOSE_MODEL=qwen3.5:4b
export GOOSE_MODE=approve
goose session --name local-ollama

Goose stores shared Desktop and CLI configuration under ~/.config/goose/config.yaml, so you only need to configure this once per machine.

Beginner usage#

Confirm Ollama is serving: curl http://localhost:11434/api/version
Confirm your model exists: ollama ls
Open your repo and launch Goose Desktop or start a CLI session
Start in approve mode if the repo has write access
Begin with a simple repo question or a small reversible edit request

Quick first run

# Terminal 1
ollama serve

# Terminal 2
cd your-project
goose session --name local-ollama

Inside Goose:

Ask: Summarize this codebase and point out the highest-risk area
Then ask: Propose one small safe refactor and tell me how to verify it

Pro usage#

Persistent CLI config#

If you want Ollama to stay your default Goose provider, keep it in ~/.config/goose/config.yaml:

GOOSE_PROVIDER: ollama
OLLAMA_HOST: http://localhost:11434
GOOSE_MODEL: qwen3.5:4b
GOOSE_MODE: approve

This gives you a durable local default for future Desktop and CLI sessions on the same machine.

Tool shim for non-tool-calling models#

Goose works best with models that support tool calling. If you want to experiment with a model that does not, Goose has an experimental Ollama tool shim:

ollama pull mistral-nemo

export GOOSE_TOOLSHIM=true
export GOOSE_TOOLSHIM_OLLAMA_MODEL=mistral-nemo
goose session --name local-toolshim

If tool use feels unreliable with local models, increase the Ollama server context window before starting Goose:

OLLAMA_CONTEXT_LENGTH=32768 ollama serve

Operational tips#

Keep Ollama and Goose on the same machine to minimize latency.
Pull models before you need them so Goose does not stall on first use.
Use smaller local models for repo exploration, summaries, and repetitive edits.
Use .gooseignore to fence secrets, generated files, or infrastructure paths you do not want Goose to touch.
If Goose misses project hints or tool context, raise the available input budget:

export GOOSE_INPUT_LIMIT=32000

Cost savings guide#

After the initial model download, local usage has no per-token API bill.
Use Ollama for repo exploration, codebase summaries, and draft edits.
Reserve cloud providers for final review or harder architectural work.
Keep a smaller local model as your default when you want cheap, fast iteration.

Privacy guide#

Local Ollama keeps prompts and repo context on your machine by default.
Goose config, transcripts, and exported sessions still live on your disk, so review them before sharing.
Prefer local extensions and local providers when you want a stricter on-device workflow.
On shared machines, avoid leaving unused cloud provider keys configured if your goal is local-only work.

Security guide#

Keep Ollama bound to 127.0.0.1; do not expose port 11434 to the public internet.
In Goose, prefer approve or smart_approve when working in sensitive repos.
Use .gooseignore to keep secrets, deployment files, and credentials out of the agent's working set.
If you ever proxy Ollama beyond localhost, put authentication and network controls in front of it.

Appendix#

Quick refs

ollama ls
ollama show qwen3.5:4b
curl http://localhost:11434/api/version
goose configure
goose session --name local-ollama

goose with ollama