Yes, agent skills work with local models, because a skill is text the agent reads and any model can follow it. You need three things: an agent that reads SKILL.md, a model that can call tools, and a context length of about 64,000 tokens. We checked all 35 repos from our weekly chart: 20 need no API key or paid service, 10 work with a catch and 5 are cloud or Claude first. The full table is below, with what to install and what still leaves your Mac.
Key facts
- Ready for local use: 20 of 35 charted repos need no API key or paid cloud service
- Works with a catch: 10 of 35 need sub-agents, extra software or optional keys
- Cloud or Claude first: 5 of 35 need paid keys or are built around Claude
- Agents to use: OpenCode, Hermes Agent, LM Studio Bionic, OpenClaw, or Claude Code through Ollama
- Model: tool-capable, with a context length of at least 64,000 tokens
- Mac: 32 GB is the sensible start; 16 GB suits small edits only
On this page
- Do agent skills work with local models?
- What you need
- Which agents run skills on a local model?
- 35 top skills checked for local use
- How to run a skill on a local model
- Why context length decides it
- What gets worse on a local model
- What still leaves your Mac?
- Frequently asked questions
Do agent skills work with local models?
Agent skills work with local models because the format is model-neutral. A skill is a folder of instructions, and the agent loads them into whatever model it is using. Nothing in the specification ties a skill to a vendor.
The catch is in the word "work". A skill written and tested on a top cloud model expects that model's ability. A smaller local model may follow the same instructions less well, skip steps or lose the thread on a long job. The author of diagram-design says as much: output quality varies by model.
What you need
- An agent that reads SKILL.md. Ollama and LM Studio's chat window run models; they do not read skills on their own. The agent is the part that does.
- A model that can call tools. Skills lean on reading files and running commands. Ollama lists suitable models under its tools filter.
- Enough context. Ollama advises at least 64,000 tokens for agents and coding tools.
- Enough memory. See our memory calculator for any model.
Which agents run skills on a local model?
Five agents cover most needs.
| Agent | Local model route | Skills | Best for |
|---|---|---|---|
| OpenCode | ollama launch opencode | Reads .agents/skills/, .claude/skills/ and its own folder | Coding, in an open-source terminal agent |
| Hermes Agent | Any local model | Native, in ~/.hermes/skills/; also writes its own | A private assistant that learns |
| LM Studio Bionic | Built in | Uses SKILL.md, including skills from Codex and Claude Code | A desktop app with no terminal |
| OpenClaw | Your choice of local model | Its own skills, shared through ClawHub | Tasks through chat apps |
| Claude Code with Ollama | ollama launch claude | Reads ~/.claude/skills/; see the note below | People who already know Claude Code |
A note on Claude Code. Skills are read by Claude Code itself rather than by the model, so they should load, but Ollama's capability list does not mention skills. Anthropic's documentation also says it does not support routing Claude Code to non-Claude models through a gateway. Treat this route as useful but unsupported.
35 top skills checked for local use
This table covers every repo to reach a weekly top 20 in our agent skills chart from 1 October to 4 November 2026. I sorted them by what each README says it needs. This is a desk check, not a test: "ready" means no key or paid service is required, not that the skill performs equally well on a small model.
Ready: no API key or paid service needed (20)
| Skill | Best rank | What it needs |
|---|---|---|
| mattpocock/skills | #1 | Plain instructions |
| ponytail | #1 | A ruleset plus hooks |
| impeccable | #1 | Plain skill; its CLI checks run with no model at all |
| ai-job-search | #1 | Runs on your machine; CVs are local LaTeX files |
| anatomy | #2 | Node 20.10 or later |
| hallmark | #2 | Node scripts; browser checks need Chrome |
| diagram-design | #3 | No build step; output quality varies by model, says the author |
| unslop | #3 | Live-site snapshots need Playwright; generated images are optional and paid |
| ai-design-skills | #4 | One markdown file |
| hyperframes | #5 | Renders locally with headless Chrome and FFmpeg |
| ECC | #6 | Plain files; its GitHub App is an optional hosted extra |
| archify | #7 | Produces a local HTML file |
| graphify | #7 | Parses code locally with no model; docs can use local Ollama |
| ppt-master | #9 | Python 3.10 or later |
| superpowers | #10 | Plain skills |
| addyosmani/agent-skills | #10 | Plain skills |
| taste-skill | #10 | Plain skill files |
| i-have-adhd | #13 | Plain instructions |
| andrej-karpathy-skills | #15 | One text file |
| ui-ux-pro-max-skill | #19 | A local Python script searches the catalogue |
Works, with a catch (10)
| Skill | Best rank | The catch |
|---|---|---|
| security-audit-skill | #1 | Runs several sub-agents through six phases, so allow for a long run |
| claude-mem | #2 | Runs a worker service alongside your agent |
| Agent-Reach | #3 | A local CLI, but it fetches from the web and some platforms read browser cookies |
| claude-plugins-official | #7 | A directory, so it depends on the plugin |
| elephant-goldfish-model | #8 | Stops the review if your agent cannot start fresh sub-agents |
| OpenMontage | #11 | Free stock media works; Ollama and LM Studio support is listed as coming soon |
| autoharness | #12 | Claude Code only, on Linux and macOS |
| scientific-agent-skills | #13 | Its desktop app uses your own keys |
| caveman | #19 | Runs locally with your own keys or login |
| hypit | #20 | Renders locally; video generation needs a hosted service, your own API or a local model |
Cloud or Claude first (5)
| Skill | Best rank | Why |
|---|---|---|
| image-blaster | #4 | Needs paid World Labs and fal keys |
| universal-modder | #7 | Needs a paid fal key for art, 3D and sound |
| Claude-Code-Game-Studios | #7 | Built for Claude Code; some agents ask for Opus or Sonnet |
| claude-seo | #18 | Its agents are set to Anthropic's Sonnet and Opus |
| knowledge-work-plugins | #18 | Built around Claude and cloud connectors |
Two patterns stand out. The simplest skills travel best: most of the ready group are text, with little or nothing to install. And the sub-agent skills are the hardest fit, because each sub-agent needs its own run through your one local model.
How to run a skill on a local model
This is the shortest route on a Mac, using OpenCode and Ollama. Commands are from Ollama's documentation.
- Install Ollama and pull a tool-capable model. See our Ollama guide and pick from models by Mac memory.
- Raise the context length to 64,000. Use the slider in the Ollama app's settings.
- Start the agent with
ollama launch opencodefrom your project folder. - Add a skill. Run
npx skills add owner/repo, or copy the folder to.agents/skills/. Read it first, using our five-point check. - Test it on a small job. Ask in plain English and see whether the agent uses the skill. Start with a single-file skill, not a big pack.
For Claude Code, the steps are the same with ollama launch claude. Our Claude Code with Ollama guide has the details and the model table.
Why context length decides it
Context length is the main trap on a local set-up, because the default can be small. Ollama sets it from your graphics memory, and below 24 GiB the default is just 4,000 tokens.
Now do the sums. The specification puts each skill at about 100 tokens at start-up, so 40 installed skills use roughly 4,000 tokens: the whole default context, before your agent's own instructions, your message or a single file. One activated skill can add several thousand more.
Two habits help. Raise the context to 64,000, accepting that this uses more memory: Ollama says one 19 GB model needs about 23 GB at that setting. And install only the skills you use, because every skill's description is in the context from the start.
| Mac memory | Usable by the model | Verdict for skills, on Claude Code with Ollama |
|---|---|---|
| 16 GB | About 12 GB | Chat and small edits only; weak agent results |
| 24 GB | About 18 GB | Works with a modest context, such as gpt-oss:20b |
| 32 GB | About 24 GB | The sensible start: gpt-oss:20b, qwen3-coder or glm-4.7-flash |
Figures are from our Claude Code with Ollama guide, based on download sizes on 22 October 2026. Buying a machine? See the best Mac or mini PC for local AI.
What gets worse on a local model
- Long, multi-stage skills. A skill with many phases asks more of the model's memory and attention. Start with short ones.
- Skills that split work into sub-agents. Elephant-goldfish-model stops its review if the agent cannot start fresh sub-agents, and security-audit-skill runs several through six phases.
- Automatic triggering. The agent must judge which skill fits. If a local model misses it, name the skill yourself.
- Speed. Expect jobs to take longer than on a cloud model. Skills that run their own scripts, such as graphify's code parsing, lose least.
A sensible split is to use a local model for private or repetitive work, and a cloud model for the hardest reasoning. Ollama Cloud sits between the two.
What still leaves your Mac?
A local model keeps your prompts on your machine. It does not make everything private, and it does not make skills safe.
- Skills can call out. Check the scripts for web addresses and API keys. Agent-Reach, for example, exists to fetch from the web.
- The agent can too. Claude Code still makes some connections of its own when pointed at Ollama, as our guide explains.
- A local model does not stop a bad skill. Its scripts still run on your computer with your access. See are agent skills safe?
For more on the wider privacy picture, read private AI.
Frequently asked questions
Do agent skills work with local models?
Yes. A skill is text that the agent reads, so it does not care which model answers. You need an agent that reads SKILL.md, a model that can call tools and enough context. Of the 35 repos in our chart, 20 need no API key or paid service, according to their READMEs.
Which local model is best for agent skills?
Use a tool-capable model with a context length of at least 64,000 tokens. For Claude Code with Ollama, qwen3-coder and glm-4.7-flash suit a 32 GB Mac, and gpt-oss:20b suits 24 GB. A skill written for Claude's best models may work less well on a smaller one, so test yours.
Can I use Claude Code skills with Ollama?
Ollama can run Claude Code against a local model with one command, ollama launch claude. Skills are read by Claude Code rather than the model, so they should load, but Ollama's capability list does not mention them and Anthropic does not support this set-up. Test your skill before relying on it.
Does LM Studio support agent skills?
LM Studio's chat app uses MCP servers and plugins. Its separate agent app, LM Studio Bionic, uses the Agent Skills format and can use compatible skills from Codex and Claude Code.
How much memory do I need to run agent skills locally?
32 GB is the sensible start. On Claude Code with Ollama, 16 GB suits chat and small edits, and 24 GB works with a modest context. The skill itself adds little: memory goes on the model and its context.
Are agent skills private on a local model?
Your prompts stay on your machine, but check the skill. Some scripts call a paid API or fetch from the web, and an agent can make connections of its own. A local model also does not make a malicious skill safe, because its scripts still run on your computer.
Read next
- What are agent skills?
- Top 20 agent skills this week
- Claude Code with Ollama on a Mac
- Run AI agents free on your Mac
- LM Studio Bionic explained
- Are agent skills safe to install?
Sources
- Agent Skills specification - checked 6 November 2026
- Claude Code, Ollama docs - checked 6 November 2026
- OpenCode, Ollama docs - checked 6 November 2026
- Context length, Ollama docs - checked 6 November 2026
- Agent skills, OpenCode docs - checked 6 November 2026
- Skills system, Hermes Agent docs - checked 6 November 2026
- Top 20 agent skills chart, each repo's README linked from its profile - checked 6 November 2026
