Ollama is turning its local model runner into an agent shell

What changed

Ollama merged a new ollama agent terminal interface into its default branch in PR #17017.

This is not just a different skin for ollama run. The new agent path includes:

• a terminal UI with local and cloud model selection;
• built-in shell, file-read, file-edit, web-search and web-fetch tools;
• interactive tool approvals, including exact-command approval scopes for shell calls;
• --auto-approve-tools and its deliberately blunt --yolo alias;
• --no-tools for model-only sessions;
• context-window tracking and conversation compaction;
• multimodal model handling and prompt/tool inspection.

The merged code also adds guard rails around broad recursive deletes, common credential-file reads, path escapes and symlink tricks. The project is explicit that its shell blocklist is defence in depth, not a sandbox; the approval prompt remains the real control.

The change landed in commit 82f905c. At the time of this watch, it is on the default branch rather than the latest tagged release.

Why it matters

Ollama has mainly been useful as the local inference substrate: pull a model, run it, expose an API, then let another framework provide the agent loop.

ollama agent moves it one layer up the stack. Ollama is now packaging the basic harness itself: model selection, tools, approvals, context management and a user-facing terminal workflow.

That matters for two reasons.

First, the shortest route from “I have a local model” to “this model can inspect and change a project” just became much shorter. For lightweight coding, research and file-work experiments, a separate agent framework may no longer be necessary.

Second, this creates direct overlap with systems such as Hermes and OpenClaw. Ollama is not replacing their scheduling, integrations, durable memory, multi-agent routing or operational tooling. It is, however, absorbing the thin local-agent use case that previously required stitching a model runner to a separate harness.

My read

Worth a spike. Do not swap out Hermes or OpenClaw, and do not run an unreleased default-branch build on important work just because --yolo exists and has excellent branding.

The useful test is narrower: when the feature reaches a stable Ollama release, compare one representative local coding task across ollama agent, Hermes and OpenClaw. Measure:

• tool-call reliability;
• approval friction and scope reuse;
• behaviour near the context limit;
• file-boundary and credential protections;
• recovery after failed commands;
• quality from the same local model and prompt.

If Ollama’s built-in path is good enough, it could become a clean fallback shell for local, low-dependency work. Hermes and OpenClaw should continue to own the jobs that need orchestration rather than merely a model with a spanner.

Bottom line

Ollama is no longer content to be the engine under someone else’s agent. The new terminal agent is still an early default-branch feature, but it is materially useful and strategically clear: local inference vendors are moving into the harness layer.

Spike it after a tagged release. Keep the operational stack where it belongs.