Base64 for AI agents

AI coding assistants and autonomous agents now paste Base64 into configs, HTTP headers and tool arguments every day. The failure mode is new: the model invents encoding instead of computing it. This page is the rulebook we follow on Base64Encoding.com.

Core principles

1. Deterministic over probabilistic — Encode/decode with a library or browser tool, never with “please Base64 this” alone.
2. Private by default — Prefer client-side tools so tokens and keys never hit an upload API (why).
3. Verify before trust — Round-trip check anything a model produced (Verify Base64).
4. One layer only — Do not Base64-wrap data that is already Base64; unwrap with nested decode if you inherit a mess.
5. Right dialect — JWTs and OAuth need Base64URL; e-mail needs MIME; certificates need PEM.
6. Watch token cost — Base64 inflates size ~33% and burns LLM context when agents inline files. Prefer URLs/resources for large binaries; use the size calculator.
7. URL hygiene — Prefer Base64URL over percent-encoding standard Base64; use URL encode/decode only for true URL components.
8. Charset honesty — European legacy text needs the right charset (guide).

What agents typically need

MCP and tool-calling note

If you expose Base64 via MCP, keep tools keyless and offline for encode/decode. Document that large images should not be returned as giant Base64 blobs inside model context—use resource URLs. Agents should call base64_encode / base64_decode tools rather than generating strings themselves.

Related reading

• ChatGPT cannot reliably decode Base64
• Private Base64 decoding
• llms.txt for AI crawlers