Guide · Agent skill

How to use the Ollama Local Models Bridge agent skill

Updated 2026-10-29 · written with AI from the project's own documentation and checked against our catalogue data

The Ollama Local Models Bridge is an agent skill, a packaged set of instructions and files that teaches an AI agent a new capability. This one, named add-ollama-tool, connects an agent running inside a NanoClaw container to Ollama, a program that runs open AI models on your own computer. This guide covers what the skill does, how to install it, what it needs, how to ask for it, and the safety settings to check before you turn on its optional management tools.

Ollama Local Models Bridgenanocoai/nanoclaw · MIT Download the skill

Quick start

1Check Ollama is running. Confirm the daemon answers at http://127.0.0.1:11434 using the curl command in the guide. If it does not, install Ollama and start it.
2Pull at least one model. Download a model with ollama pull, such as gemma3:1b for a small model or llama3.2 for general use.
3Download and unzip the skill. Get the ZIP from this page and unzip it into ~/.claude/skills/ or a project's .claude/skills/ folder, or upload it under Settings, Capabilities, Skills in the Claude apps.
4Apply the skill to your NanoClaw repo. Follow the SKILL.md phases to copy the files, register the ollama server in index.ts, and forward the host environment variables.
5Run the validation checks. Run the build, the type check, the two wiring and registration tests, and the container build. All must pass before you use the tools.
6Ask for a local model task. Make a normal request that names the model or task. Enable OLLAMA_ADMIN_TOOLS only if you want the agent to manage the model library.

What does the add-ollama-tool skill make the agent do?

Once installed, the agent can use a small Ollama server that exposes local models as tools. Claude stays the orchestrator, but it can hand some work to a model running on your machine. That can suit simple tasks where a cheaper or private local model is enough, or where you want to compare a local model's output against Claude.

The core tools are always available: ollama_list_models lists the models you have installed, with name, size and family, and ollama_generate sends a prompt to a chosen model and returns its response. An optional setting adds four management tools: ollama_pull_model downloads a model, ollama_delete_model removes one to free disk space, ollama_show_model displays details such as the modelfile, parameters and architecture, and ollama_list_running shows which models are loaded in memory with their memory usage and processor type.

Core (always on): ollama_list_models, ollama_generate
Management (opt-in with OLLAMA_ADMIN_TOOLS=true): ollama_pull_model, ollama_delete_model, ollama_show_model, ollama_list_running

How do you install the add-ollama-tool skill?

Download the skill ZIP from the page where you found this guide, then unzip it. Put the folder in one of these places: ~/.claude/skills/ to make it available to you across projects, or a project's .claude/skills/ folder to limit it to that project.

If you use the Claude apps instead, open Settings, go to Capabilities, then Skills, and upload the skill there.

The skill folder contains SKILL.md, which holds the full instructions, plus the server and test files ollama-mcp-stdio.ts, ollama-env.ts, ollama-registration.test.ts, ollama-wiring.test.ts, and REMOVE.md for uninstalling.

What do you need before the skill works?

The skill is built for a NanoClaw repository that has the container agent-runner and host source trees. Ollama must be installed on the host and its daemon must be running, by default at http://127.0.0.1:11434. Without a running daemon, the tools will fail. You also need at least one model pulled locally.

Ollama is keyless, so there are no credentials to set. The only configuration is the daemon's base URL, which the OLLAMA_HOST variable can override.

Ollama installed on the host, daemon running
At least one Ollama model pulled locally
NanoClaw repository with container agent-runner and host source trees
Bun (container) and Node test tooling as used by the repo

curl -s http://127.0.0.1:11434/api/tags | head

How do you apply the skill to a NanoClaw repository?

The skill copies its server and test files into both the container (Bun) tree and the host (Node) tree, then adds an ollama entry to the mcpServers object in container/agent-runner/src/index.ts. It also imports and spreads an ollamaEnv() helper in src/container-runner.ts so OLLAMA_HOST and OLLAMA_ADMIN_TOOLS reach the container. If container/agent-runner/src/ollama-mcp-stdio.ts already exists, the skill says to skip to the configuration step.

The skill also edits the shared [OLLAMA] log branch in src/drivers/docker-driver.ts. That block is shared with other local-model tools, so only the [OLLAMA] branch should change.

Copy the four files into both trees
Register the ollama server in index.ts
Forward OLLAMA_HOST and OLLAMA_ADMIN_TOOLS in container-runner.ts
Surface [OLLAMA] stderr lines at info level
Add the env-var stubs to .env.example
Run the build, type check, tests and container build

S=.claude/skills/add-ollama-tool
cp $S/ollama-mcp-stdio.ts container/agent-runner/src/ollama-mcp-stdio.ts
cp $S/ollama-registration.test.ts container/agent-runner/src/ollama-registration.test.ts
cp $S/ollama-env.ts src/ollama-env.ts
cp $S/ollama-wiring.test.ts src/ollama-wiring.test.ts

How do you ask the agent to use the Ollama tools?

You trigger the tools with ordinary requests. Name the local model or the job you want done, and the agent can pick the matching tool. For example, you can ask which local models are installed, or ask a named model to draft or summarise something while Claude coordinates the rest of the work.

Management tools only appear when the admin setting is on. If you ask the agent to pull, delete or inspect a model and those tools are missing, check that OLLAMA_ADMIN_TOOLS=true is set on the host.

Which Ollama models do I have installed?
Ask the local model to summarise this text.
Show me details for the model I have loaded.
Which models are currently using memory?

What are the safety notes for scripts and network access?

This skill runs scripts and uses the network. The Ollama daemon is reached over HTTP, so anyone who can reach that daemon can use it. Avoid enabling it on shared or multi-user machines where other people could connect to the local Ollama daemon.

Enabling OLLAMA_ADMIN_TOOLS lets the agent delete models and pull large downloads. Only turn it on if you are comfortable with the agent managing disk space, and check what it plans to do.

Runs scripts: yes
Uses network: yes
Keep OLLAMA_ADMIN_TOOLS unset unless you want library management
Do not expose the Ollama daemon on shared machines

How do you remove the skill?

The REMOVE.md file in the skill folder describes a removal that is safe to run even if some steps were never applied. It deletes the four copied files from both trees, removes the ollama entry from mcpServers, and reverts the host-side edits, including the ollamaEnv import, the spread in composeSessionSpec, and the [OLLAMA] branch in the stderr handler. Follow REMOVE.md rather than deleting files by hand, so the shared log block is restored correctly.

What people use it for

Ask a small local model to draft or summarise text while Claude handles planning.
Compare a local model's answer with Claude's answer on a simple task.
Keep prompts that should stay on your own hardware on a local model.
Check which models are installed and how much memory the loaded ones use.
Clean up unused models to free disk space, if admin tools are enabled.

Frequently asked questions

Is the add-ollama-tool skill free?

The skill is released under the MIT licence. This page does not state pricing for Ollama itself, so check Ollama's own site for its terms before you rely on it.

What do I need to run the Ollama skill?

You need Ollama installed on the host with its daemon running, at least one model pulled locally, and a NanoClaw repository with the container agent-runner and host source trees. The tools will fail without the running daemon.

Why do the Ollama tools fail with an error?

The most common cause is that the Ollama daemon is not running. Check that http://127.0.0.1:11434/api/tags answers with curl. If you set OLLAMA_HOST, make sure it points to the right address.

How do I turn on model pull and delete?

Set OLLAMA_ADMIN_TOOLS=true on the host to enable pull, delete, show and list-running. Leave it unset if you only want listing and generation, because the admin tools can delete models and start large downloads.

Does the skill send my prompts to the cloud?

The skill sends prompts to the Ollama daemon you run, which the source describes as running locally and keyless. Prompts go to whichever model you choose, so check the model and daemon setup if privacy matters to you.

How does this compare with other NanoClaw local model tools?

NanoClaw has other local-model skills, such as add-atomic-chat-tool, which uses its own [ATOMIC] log prefix. Both edit the same stderr logger block, so they can be used together. This guide does not compare their output quality.

New agent skills, repos and AI tools — every week

The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.

Our verdict

Use this skill if you run NanoClaw, already have Ollama working on your machine, and want some tasks handled by local models while Claude coordinates. Pick something else if Ollama is not installed, if the machine is shared with other users, or if you do not want an agent that can delete models and pull large downloads.

Download the skill