If you use Claude Code for anything non-trivial, you've probably had the moment where it does something and you think: what exactly just happened there? Both agents-workbook and Claude-Sniff answer that by sitting between you and the model as a local proxy — but they answer a different version of the question, and they're not really competing for the same job.
The short version: agents-workbook shows you the model's reasoning — what it considered and rejected — on a live dashboard, at the cost of doubling your API calls per turn. Claude-Sniff shows you the actual request Claude Code sends to AWS Bedrock, and lets you edit it before it goes out — but only if you're running Claude Code on Bedrock specifically.
agents-workbook — watch the model think, across providers
agents-workbook is a local proxy that lets Claude Code or Codex stream their reasoning to a dashboard in real time, including the alternatives the model considered and rejected along the way, plus a running cost analysis. It's open source (Apache 2.0), deploys with a simple Docker Compose setup, and — because it works with both Claude Code and Codex — isn't locked to a single backend.
Forces: real-time reasoning stream, not a post-hoc log; captures rejected alternatives, which most tools throw away; 100% local, nothing sent to a third party; works with more than one agent CLI.
Limites: doubles your API calls per turn to generate the reasoning stream — a real token cost, not a rounding error; no team dashboard or cross-session history, it's built for one developer on one machine; scope is deliberately narrow.
Claude-Sniff — intercept and edit the raw Bedrock request
Claude-Sniff is a local proxy that lets you inspect and edit Claude Code's requests to AWS Bedrock before they're sent — intercepting and modifying them before SigV4 signing, with a web UI for real-time inspection and full JSONL logging. It also tracks cost, computed from request metadata and regional pricing, which is otherwise annoyingly opaque on Bedrock.
Forces: genuine request-level control — you can actually change what gets sent, not just observe it; built-in cost tracking specific to Bedrock's regional pricing; full JSONL logs for later inspection.
Limites: only useful if you run Claude Code on AWS Bedrock — a narrow setup by itself; only intercepts requests, not responses, and doesn't mask request bodies; very early and effectively unproven (0 stars, 0 forks, 2 commits at time of writing).
| agents-workbook | Claude-Sniff | |
|---|---|---|
| What you see | Model reasoning + rejected alternatives | Raw outgoing API request, editable |
| Backend requirement | Claude Code or Codex, any backend | Claude Code on AWS Bedrock specifically |
| Token cost | Doubles API calls per turn | None beyond normal usage |
| Can you change what's sent? | No, observation only | Yes, edits happen before signing |
| Price | Free, open source | Free, open source |
Pick agents-workbook if you want to understand the model's decision-making process itself — especially the paths it didn't take — and you don't mind paying extra tokens for the visibility. Pick Claude-Sniff if you specifically run Claude Code through AWS Bedrock and need to inspect or tweak the exact request before it leaves your machine — for debugging a Bedrock-specific issue that a generic reasoning dashboard won't show you. They solve adjacent but distinct problems, and given both are free and local, there's nothing stopping you from running whichever matches the debugging session you're actually in.