AutoWeight

Model weight editing

Edit model weights.
Verify the result.

Use Codex CLI or Claude Code to propose edits to a local Hugging Face model. AutoWeight tests each candidate against the behaviors you want to change and preserve.

Connect your CLI Try the offline demo

Python 3.11+ · macOS and Linux · Runs locally

Offline demo preview

Scripted run

Repair run

Tiny Hugging Face model

Restore raw JSON responses while preserving queue extraction.

Confirmation results

Accepted
BehaviorBaseEdited
Correct raw JSON0/3232/32
Queue extraction32/3232/32

Exported as a Transformers checkpoint

Results from the included CPU demo. Run it locally to create the model and reproduce the export.

Agent setup

Connect Codex CLI or Claude Code

Your coding agent proposes weight edits through MCP. AutoWeight applies and scores them locally, then exports a checkpoint if the selected edit passes confirmation.

01 Prepare your inputs and install AutoWeight

Use an existing input directory with a Hugging Face causal language model in base/ and independent development.jsonl and confirmation.jsonl files. Replace ./repair-inputs with that directory.

Terminal · setup
uv tool install 'auto-weight[hf,mcp]'
auto-weight setup-runtime --input-root ./repair-inputs

Model and evaluation file format

Include the model’s safetensors weights, config, and tokenizer files. Put at least 32 distinct labeled examples per group in each JSONL split. Give the behavior to change the group fix and give each behavior to preserve its own group. Use different prompts in the confirmation set and leave those examples for the runner.

{"group":"fix","prompt":"Return raw JSON for priority high.","check":{"type":"json","expected":{"priority":"high"}}}
{"group":"queue","prompt":"Name the queue for invoice questions.","check":{"type":"exact","expected":"billing"}}

02 Register the server with your CLI

Run the commands for your installed client from the project you work in. Then use /mcp in a new session to check for auto-weight.

Codex CLI
codex mcp add auto-weight -- \
  "$(command -v auto-weight-mcp)"
codex

Claude Code
claude mcp add --transport stdio auto-weight -- \
  "$(command -v auto-weight-mcp)"
claude

Codex MCP docs · Claude Code MCP docs

03 Tell the agent what to edit

Example prompt
Use AutoWeight to remove Markdown fences from JSON responses in the model at base. Improve the fix group and preserve queue extraction. Use development.jsonl to try edits and reserve confirmation.jsonl for one final check. Export the edited checkpoint if it passes.

Workflow

From proposed edit to verified checkpoint

  1. 01

    Define the checks

    Bring a local Hugging Face checkpoint and labeled examples for the behavior to fix and preserve.

  2. 02

    Test targeted edits

    Ask Codex CLI or Claude Code for an edit. AutoWeight applies it to a temporary checkpoint and evaluates it on development examples.

  3. 03

    Confirm and export

    Score the selected edit on held-out cases. A passing run exports the evaluated model and its report.

Get started

Run the offline demo

The demo builds a tiny model on CPU and works offline after installation. Use Python 3.11 or newer on macOS or Linux.

View the package on PyPI

Terminal · uv
uv tool install 'auto-weight[hf,mcp]'
auto-weight demo

Install with pip 
python3 -m venv .venv
source .venv/bin/activate
python -m pip install 'auto-weight[hf,mcp]'
auto-weight demo

Research

What holds up in the experiments?

Read five agent repair cases and eight weight-space studies, with local replay results distinguished from the published benchmarks.

Read the research note