A coding agent that redacts your secrets before the LLM ever sees them

What it does

Baize is an open-source coding-agent CLI for “Vibe Coding” and data analysis, with a privacy layer as its headline feature. Before anything is sent to the LLM, sensitive data — phone numbers, emails, ID cards, API keys, JWTs, database connection strings, and 30-odd other categories — is replaced with cryptographically random placeholders like «PZ_PHONE_a3f8b2c1d4e5f6a7», then restored in the response so you only ever see the original text. It talks to DeepSeek, any OpenAI-compatible endpoint (Qwen, Kimi, GLM, OpenAI), or a local Ollama, and runs entirely on your machine with no remote database.

The interesting bit

The redaction isn’t just in transit: session files, compressed transcripts, and audit logs are scrubbed before hitting disk, so plaintext PII never lands anywhere. The placeholders use 64-bit random tokens and die when you switch sessions, which rules out enumeration, forgery, and replay — a genuinely thoughtful design for a problem most agents ignore entirely.

Key highlights

• Bidirectional, reversible redaction across 30+ PII categories, with Standard and Strict modes
• Multi-backend support: DeepSeek, OpenAI-compatible endpoints, and local Ollama, switchable in one config field
• Full agent toolkit: bash execution, file editing, grep/glob, web search, subagents, skills, hooks, and MCP
• Two-level context compression (tool-result truncation plus LLM summarization) for long conversations
• Secure sandboxing: command allowlists, path-escape detection, and dangerous-command interception
• Nine-language interaction, switchable by just saying the language name

Caveats

• The SPSS integration is Windows-centric — Linux/macOS only get read-only “file mode” with no statistical analysis, and requires a licensed local IBM SPSS install anyway.
• SQL Server support needs system-level ODBC Driver 18 installed separately.
• The repo is tagged JavaScript but the tool itself is Python 3.10+; the language label is a bit of a red herring.

Verdict

Worth a look if you want a coding agent that treats your API keys and customer data as things the model shouldn’t see. If you’re happy piping raw context to a hosted model and don’t touch SPSS, it’s a harder sell.