[FEATURE] Local Anthropic specialist models (OCR first), ideally open-weight, that Claude Code can use as delegates

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Some recurring jobs are narrow, high-volume and well-bounded: OCR, transcription, structured extraction. For these jobs Anthropic currently offers only cloud models, so every page or file counts against usage limits or API costs. For high-volume work, the limits can run out long before the job is done.

A local layer around Claude already exists. Ollama and LM Studio expose Anthropic-compatible endpoints, and the community has built MCP servers, skills and subagents that let Claude Code delegate such jobs to local models. But every model in that layer comes from another company. Anthropic has no model that people can run on their own machines, so it is absent from the layer where many people now build their own tools, products and data pipelines.

Small open models are the seeds of that ecosystem: people build on what they can download, adapt and run. At the moment those seeds come from Anthropic's competitors. The current workarounds are also too technical for most users, who have to choose a runtime and a model and write the integration themselves.

Proposed Solution

The ideal form, in my view: release small, task-specialised Anthropic models as open weights, for example on Hugging Face, and give Claude Code first-party support for using them. OCR would be a natural first candidate. The models should be sized for ordinary hardware: Apple Silicon Macs with 24 GB of unified memory, or 16 GB where the task allows.

First-party support in Claude Code would mean:

• installing and updating a specialist model from within Claude Code, without users having to assemble a runtime, a model and an MCP server themselves;
• exposing installed specialists to Claude as local tools, so that Claude, or a project coordinator, can route bounded jobs to them (OCR a batch of page images, extract fields, transcribe recordings).

Because the weights would be open, the same models could also be used directly in people's own tools and pipelines, outside Claude. If open weights are not acceptable to Anthropic, models that run locally only inside Claude's harness would still solve the usage and integration problems, though not the wider ecosystem one.

One design point decides whether delegation actually saves anything: the specialist should mark uncertain output instead of guessing. Claude then reviews only the flagged items plus spot checks. If Claude had to re-check everything, delegation would simply duplicate the work and the usage. For OCR, this also means faithful transcription: reproducing what is printed, including historical spelling, rather than silently "correcting" it.

In this form I cannot see who would lose out:

• users get capable local models, much lower usage for high-volume work, and a way to keep material on their own machines when it must not leave them;
• developers and researchers can build products and pipelines on Anthropic models, and produce high-quality outputs and data at a scale that usage limits would never allow;
• Anthropic keeps its models at the centre of local pipelines instead of leaving that layer to its competitors, and moves narrow, high-volume work off its servers onto hardware users already own, where a small specialist needs far less computation per task than a frontier model.

Alternative Solutions

Cloud models only: Haiku is cheaper than Sonnet or Opus, but it still runs in the cloud and counts against the same usage limits, which is exactly the problem for high-volume work.

Non-Anthropic local models: my own OCR tool already relies on non-Anthropic local models and tools, and many people connect such models to Claude Code through Ollama or LM Studio and community MCP servers or skills. This works, which shows that the plumbing largely exists. But the local half of the pipeline is then built on other companies' models, quality and faithfulness vary from model to model, Claude has no first-party knowledge of the delegate's strengths and failure modes, and the setup remains out of reach for most users.

Priority

High - Significant impact on productivity

Feature Category

API and model interactions

Use Case Example

I am working on eighteenth-century music sources, and I build my research tools with Claude Code. One of them is an OCR tool for English printed sources from 1750 to 1800, such as newspapers and catalogues. It runs entirely on a Mac mini.

1. Today Claude is central to building the tool but has no part in running it, because the OCR itself has to rely on non-Anthropic local models and tools.
2. With an open-weight Anthropic OCR model, I would run bulk OCR directly in my tool, on my own machine, without spending usage.
3. In a Claude Code session or project, Claude could send a batch of pages to the same local model and then review only the lines it flagged as uncertain, plus a sample for quality control.
4. The heavy work would run locally, Claude would do the part that needs a frontier model, and the whole pipeline would stay within the Anthropic ecosystem.

Step 3 would also work with a model that runs only inside Claude's harness; step 2 needs open weights.

Additional Context

This follows #91269 (Project/Contractor Mode). About two weeks later the Projects redesign was announced, which I was very glad to see. The announcement notes that projects can reach usage limits faster, and that threads will soon be able to run on the user's own machine. Once they do, a specialist model on the same machine would be a natural tool for them: the coordinator decides, and bounded work goes to the cheapest executor that can do it well.

On safety: I understand Anthropic's caution about open weights, and I broadly share it. This request is deliberately narrow. Small open-weight models that can do these tasks are already available from several companies, so a small Anthropic specialist would not move the capability frontier; what it would change is whether Anthropic is present in this layer at all. I believe its marginal risk would be low. And since a local model has to live on users' machines whichever form it takes, I think publishing it openly would add little risk beyond a harness-only release, while adding much more value for the wider ecosystem.

Model releases are of course decided outside this repository. I am filing this here because Claude Code is where the integration would live, and where the demand is already visible in community tools built for local delegation.

OCR is only the first example; the same pattern would apply to transcription, layout analysis and structured extraction.