Ollama models

Ollama is Local Agent's only implemented inference runtime. The model picker is driven by Ollama's live inventory, including installed local weights and available Ollama Cloud aliases.

Open the inventory

Press ctrl+o or run:

/model

The picker groups entries by execution location:

• Local — model weights available on this Ollama instance.
• Cloud — aliases reserved for manual selection; under local-only policy they require exact conversation consent.
• Remote — models exposed by a non-Cloud remote host; these are visible but not selectable.

Any row can also be unavailable when blocked by runtime capability, context metadata, privacy policy, or the local memory guard.

Use d for runtime and capability details, a to open the cancellable pull form, and r to refresh the inventory.

Automatic routing and pins

The TUI starts with availability-aware automatic routing. It considers admitted local models only.

/model qwen3.5:4b   switch to and pin one model
/model auto         release the pin and resume automatic routing
/model list         print the current admitted inventory

Choosing a verified local model from the picker also pins it and remembers the choice across process restarts. /model auto clears that saved choice. An explicit startup --model flag or agent-profile model takes precedence, while Cloud consent remains limited to the current conversation and is never restored implicitly.

Local and Cloud are different boundaries

Cloud models can remain visible when privacy.local_only: true. Selecting one is a deliberate exception for the current conversation:

1. Choose the Cloud entry.
2. Review the destination and privacy boundary.
3. Confirm the exact model for this conversation.

Cloud models are never chosen by automatic routing. Consent does not change the saved local-only setting or authorize a different Cloud model.

Context windows

Local Agent reconciles the live model's native context with the configured local request limit:

• Local requests use the lower of the configured ollama.num_ctx and the model's reported native maximum.
• Cloud requests omit the local num_ctx option and display the Cloud model's reported native maximum.
• Unknown Cloud context limits fail closed instead of displaying a guessed number.

This keeps session statistics aligned with the request that was actually sent.

num_ctx is a request and KV-cache allocation, not a promise that the whole native model maximum is practical on the machine. For example, a model can report a much larger native limit while Local Agent correctly sends the smaller configured value. Start at 16384 for the documented compact 2B/4B tiers and raise it only after accounting for model weights, the embedding model, and other local processes. Pair small models with a lazy MCPHub profile so the first provider request does not include every downstream tool schema.

Suggested local tiers

Artifact sizes vary by Ollama build and quantization. The shipped guard is tuned for a 16 GB Apple-silicon machine, not measured free memory.

Model	Intended use	Auto routing
qwen3.5:0.8b	Short answers and lightweight classification	Eligible
qwen3.5:2b	Compact interactive work and modest tool chains	Eligible
phi4-mini:latest	Alternative compact reasoning profile	Exclusive
qwen3.5:4b	Coding, debugging, review, and multi-step tools	Eligible
qwen3.5:9b	Deeper manual profile	Exclusive
gemma4:e2b	Multimodal compact profile	Exclusive
ornith:latest	Manual agentic coding and verification profile	Exclusive

The shipped Phi profile is manual-only because its advertised Ollama metadata has not been backed by behavioral tool-use verification in this harness. Its size is not the reason for the restriction.

Large-model admission can be overridden with LOCAL_AGENT_ALLOW_LARGE_MODELS=1, but the override does not add memory isolation or make an oversized model safe for the machine.

Troubleshooting

The first request already fills the context window

Check the active ollama.num_ctx and model with /model list. A large native context advertised by Ollama does not override the configured request limit. Then reduce the recurring MCP schema surface: use MCPHub gateway mode with expose: lazy, pin one useful first-call tool like bob__context (canonical name as of MCPHub 0.3.0) with tool_schema_budget: "8KB", and restart the agent after syncing the change. The MCP guide shows the complete public configuration.

Local Agent estimates prompts before dispatch and may compact tool definitions for one pressured turn. If it still cannot keep the prompt below the admission threshold, it stops instead of sending an overfilled request. Lower the amount of loaded context or use a model and num_ctx combination that fits available memory; raising a model's advertised maximum does not solve an oversized first prompt.

If the picker is empty or stale:

ollama list
curl http://localhost:11434/api/tags

Then press r in the picker. An offline refresh keeps only a previously verified safe local model; it never reclassifies a Cloud alias as local.