REALSTOP / CLAUDE CODE / RESEARCH PREVIEW

Less “Want me to?”
More doing.

A local Stop hook for Claude Code that catches work handed back to you—and lets real approval stops through.

Apache-2.0 · Local judge on Apple silicon · Phrase-rule fallback
Approval enforcement stays in the harness. Codex integration is not supported in this release.

What it does—in three stops

Actual hook decisions on synthetic fixtures using the 0.1.1 phrase rules. Not a live task-completion demo or model benchmark. Static explanation.

01 / HANDED BACK

“Want me to add the requested tests?”

RealStop can block the stop and return feedback so the agent continues.

02 / APPROVAL NEEDED

“May I deploy to production?”

The stop goes through. You decide whether the gated action is authorized.

03 / CLAIMED COMPLETE

“Implemented. All three tests pass.”

The turn can end. RealStop reads the ending; it cannot verify that the work was done.

Install the hook

Local judge · Apple silicon

Install and register RealStop, then keep the model server running in a separate terminal.

uv tool install realstop
realstop setup --judge local
realstop serve

CLI: Python 3.9+. Local server: Apple silicon and Python 3.10+. Setup downloads about 1.5 GB of model files; the server runtime needs additional disk space. Until the server answers, phrase rules decide.

No model? Use phrase rules.

uv tool install realstop
realstop setup --judge rules

No model download and no external judge call. Lower setup cost, with different measured trade-offs.

Check with realstop status. Pause with realstop pause; remove with realstop uninstall. In Claude Code, type exactly realstop off to bypass it.

Full installation and troubleshooting ↗

Evidence, with the trade-offs visible

67.5%

Hand-backs caught

15.8%

Right stops wrongly blocked

27 ms

Median warm model time

Default local setting on 864 development-used test turn ends from one developer. Timing measured on Apple M5 Pro, macOS 27.2; full-hook median 186 ms. These are reported offline results, not a fresh independent holdout or cold-start measurement.

What the live pilots showed

Fewer closing offers. They were too small to establish that agents finish more work.

What it still misses

False “done” claims, unfamiliar phrasing, and some short replies. The model is weak on the public synthetic benchmark.

Read the model card ↗ Inspect RealStopBench ↗

Questions developers ask

Does RealStop work with Codex?

Not in this release. The published installer and transcript handling target Claude Code. Codex sessions appearing in the research data do not establish runtime support.

Does my code leave my computer?

In local and rules modes, RealStop sends no prompt or reply text to an external judge. Claude Code still communicates with its own provider. Optional cloud judging sends unredacted request/reply excerpts, tool names and learned examples to that provider; read the repository's privacy section before enabling it.

Can it bypass approvals or prove a task is complete?

No. RealStop is not a permission boundary or a verifier. Keep harness permissions enabled. Its pattern-based approval handling and decisions can be wrong.

Can I adjust how often it intervenes?

Use realstop strength light, moderate, or hard. Higher strength catches more hand-backs but also blocks more right stops. The model card reports both rates.

Less “Want me to?”More doing.