Self-hosted, Self-Improving, Federated Everything App — dev management, agent orchestration, persistent mind, digital twin, second brain, and creative pipeline from a single dashboard over Tailscale to port the user into the OS
Claude Ollama — run CoS agent tasks on a local model
Claude Ollama lets a user whose only LLM is Ollama / LM Studio run file-writing CoS agent tasks — create/edit files, run tests, ship a PR — by running the Claude Code harness on top of a local model. It ships in two variants:
- Claude Ollama (local model) — a
cliprovider (claude --print), headless. - Claude Ollama TUI (local model) — a
tuiprovider (claude --dangerously-skip-permissions), the interactive harness CoS agents drive in TUI mode.
Since #7561 this is a preset example. Claude Ollama is what the composed model calls
claude.cli@ollama/claude.tui@ollama: the Claude Code harness pointed at the Ollama service instance. The shipped records below are presets of that pair, derived from the Ollama service where the boot backfill can prove it, so editing the service's endpoint reaches both. Any other enabled harness × service pair is composed the same way — see AI_PROVIDERS.md. The rest of this page still describes how the pair works and how to set it up.
Why a special provider?
PortOS has two provider classes:
cli/tui(claude, codex, agy) — spawned as a child process in the agent worktree. The CLI owns the Read/Write/Edit/Bash tool harness, so it can write files. CoS agent tasks run only on these.api(Ollama, LM Studio, NVIDIA NIM) — an HTTP request that returns plain text. No harness, no file writes. Great for planning/vision/ask, but a task that needs to edit files produces a transcript and writes nothing.
Claude Code's harness is independent of which model serves tokens. Point the
claude CLI at a local model via ANTHROPIC_BASE_URL and you keep the entire
file-editing harness while Ollama does the generation. Claude Ollama is exactly
that: a claude CLI/TUI provider whose envVars route Claude Code to your Ollama
daemon, with its model list pulled live from Ollama.
Setup
- Install the
claudeCLI on the PortOS host —npm i -g @anthropic-ai/claude-code(the harness runs locally; only token generation is remote). - Run Ollama ≥ 0.14 (locally or on a remote box) — it natively serves the Anthropic Messages API, so no proxy is needed. (LM Studio / vLLM / older Ollama: put LiteLLM in front to translate Anthropic ↔ OpenAI.)
- Pick a variant — after
./update.sh(or a server restart), both Claude Ollama (local model) (CLI) and Claude Ollama TUI (local model) appear on the AI Providers page, disabled by default. Enable the one you want. EditANTHROPIC_BASE_URLif your Ollama isn't onlocalhost:11434(e.g.http://<remote-host>:11434). - Refresh models — hit the provider's Refresh Models button. PortOS pulls the installed Ollama models and keeps only the tool-use-capable ones (the harness depends on reliable tool-calling), then pick a default. This works for both the CLI and TUI variant.
- Run a task — set Claude Ollama as the active provider (or pin it per task) and create a CoS agent task. It now writes files and ships a PR backed by your local model.
Both providers are enabled: false by default — adopting them doesn't change your
active provider until you choose one.
Coding-task restrictions
- The task form only offers coding providers (CLI/TUI). Raw Ollama / LM Studio /
kimi (
api) providers are not file-writing runners and are excluded. If your only enabled providers areapi-type, the form shows an advisory pointing you here. - A server-side guard in
agentProviderResolutionrejects anapiprovider that reaches the spawn path (via a task pin or the fallback chain) with a clear error, rather than spawning a child process that can't write files. - Claude Ollama's model list is filtered to tool-use-capable models (Qwen 2.5/3, Llama 3.1+, Mistral/Mixtral, Cohere Command, GLM-4, Granite 3, gpt-oss, Hermes, …).
⚠️ Tool-use caveat
Small generic models (Mistral 7B, base Llama, Phi) "run" but fail ~35–85% of tool
calls and silently stop editing files mid-task. Use a tool-capable model and treat
Q4_K_M as the practical quant floor. The model-list filter steers you toward
reliable families, but a model that reports the tools capability can still be weak
in practice — prefer ≥7B instruction-tuned models from the families above.
See docs/examples/claude-code-ollama-settings.json
for the equivalent global ~/.claude/settings.json approach.