Documentation

Using Waymaker with a Local Model

You can run an open-weight model on your own Mac and connect it to Waymaker's MCP server, so the model's reads and writes land in your workspace while the model itself never leaves your machine. This page is a recipe for people who already run Ollama and a coding agent (Claude Code or OpenCode). It is not a Waymaker feature: Waymaker does not install, run, or support the model.

What the model does stays on your Mac. What the tools do does not: every tool call is an ordinary request to mcp.waymakerone.com, made with your key, and the results are sent back to the model on your Mac.

What we measured

Tested on one Apple M5 Pro with 48 GB, Ollama 0.40.1, one run per task. Treat it as a signal, not a benchmark.

Model	Memory in use	Speed	Result
`gpt-oss:20b`	about 12 GB	about 60 tokens/s	Finished all five test tasks in OpenCode (a vault note, a vault search, creating a Spark, a bug fix, a three-file rename)
`qwen3-coder:30b-a3b`	20 to 22 GB	71 to 76 tokens/s	Finished the coding tasks in Claude Code; completed two of five in OpenCode, and was unreliable at calling tools

Not tested: any Mac with less than 48 GB. gpt-oss:20b should fit in 24 GB, tightly, but we have not measured it.

Before you start

An API key with the permissions you want the model to have. Give it :read only if all you want is search and summaries. See Connecting an MCP Client.
A model pulled in Ollama, and about 1.5 times the model's size free on disk.
A context of at least 64k tokens. This is the setting that decides whether it works: at Ollama's default of 32k, Claude Code compacted its own conversation halfway through a task and stopped without finishing.

Set the context explicitly

Create a model alias that carries the larger context. It is a tiny file that shares the weights:

ollama pull gpt-oss:20b
printf 'FROM gpt-oss:20b\nPARAMETER num_ctx 65536\n' > Modelfile
ollama create gpt-oss-20b-64k -f Modelfile

Claude Code

Ollama serves an Anthropic-compatible endpoint, so Claude Code runs against it unchanged:

export ANTHROPIC_BASE_URL=http://localhost:11434
export ANTHROPIC_AUTH_TOKEN=ollama
export ANTHROPIC_API_KEY=
export CLAUDE_CODE_MAX_CONTEXT_TOKENS=65536
export ANTHROPIC_DEFAULT_HAIKU_MODEL=gpt-oss-20b-64k
claude --model gpt-oss-20b-64k

Claude Code will warn that the model is not in its catalogue and will otherwise assume a 200k window, which is why CLAUDE_CODE_MAX_CONTEXT_TOKENS matters. Its own system prompt and tools take 15 to 20k of your 64k before any work starts.

Add the Waymaker server with an .mcp.json as shown in Connecting an MCP Client. Use --strict-mcp-config if you want it to see only that server.

OpenCode

{
  "provider": {
    "ollama": {
      "npm": "@ai-sdk/openai-compatible",
      "options": { "baseURL": "http://localhost:11434/v1" },
      "models": { "gpt-oss:20b": { "tool_call": true } }
    }
  },
  "mcp": {
    "waymaker": {
      "type": "remote",
      "url": "https://mcp.waymakerone.com/mcp",
      "headers": { "Authorization": "Bearer {env:WAYMAKER_API_KEY}" },
      "oauth": false
    }
  }
}

Keep the key in the environment (WAYMAKER_API_KEY) rather than in the file.

What to watch for

The tool catalogue is large. Waymaker exposes a large catalogue of tools, and its full schema is far bigger than a local model's context can hold. OpenCode copes by hiding tools behind a search step, and that step was where both models made most of their mistakes. There is no smaller, curated tool set for local models today. Restrict the tools your client offers to the few you need, and give the key only the permissions the task requires.
A small model can over-reach. In one test a model asked only to search went on to create a note. Use a read-only key for anything you do not want written, and review before approving writes.
Expect slower, plainer answers than a frontier model. Choose the local model for routine work, not for anything where a wrong answer is expensive.
Memory. The 30B coding model took a 48 GB Mac to 18 percent free memory. Quit heavy apps first, and let Ollama unload the model when idle (OLLAMA_KEEP_ALIVE=5m).
Disk. Ollama 0.40 kept a second copy of one model for a different runner. Check ~/.ollama/models if space matters.
Credits. The local model's own inference is not metered, because Waymaker never sees it. Waymaker tool calls are the same requests any client makes, and are subject to your key's limits.