MCP Payload Optimization — New Subpolicies & Configuration

What's New

The MCP Payload Optimization policy is now an umbrella policy that groups multiple independent optimizations together. Its original behavior is fully preserved and continues to work exactly as before without changes to your deployment. On top of it, we have appended two new optimizations as additional subpolicies within the same umbrella policy:

Subpolicy	What it does	Default
Smart Response Trimming (new)	Uses an LLM proxy to read what the caller actually asked for and trim the response down to just the relevant parts.	Off
Compress Repeated Structures (new)	Rewrites repetitive JSON (such as arrays of similar records) into TOON, a compact format that carries the same data in fewer tokens.	Off

Configuration

1. MCP Payload Optimization (on by default)

Strips out content that costs tokens but doesn't help the model. Four switches:

Setting	What it does	Default
HTML to Markdown	Converts HTML content to cleaner Markdown	On
Strip Base64	Removes embedded base64 data like images and PDFs	On
Remove Nulls	Drops empty (null) fields	Off
Collapse Whitespace	Trims extra spaces, tabs, and line breaks	On

Supporting options:

Min Response Size — responses smaller than this (default 500 bytes) are left alone.
Per-Tool Overrides — change these four settings for specific tools (for example, allow one tool to keep its base64 images).

2. Smart Response Trimming (off by default — new)

Uses an LLM through a user-provided LLM proxy to read what the caller actually asked for and trim the response down to just the relevant parts. Because it relies on an user-configured LLM proxy, turning it on reveals a few required settings:

Setting	Purpose
LLM Proxy URL	The endpoint used to do the trimming
Model-Based Routing / LLM Model	Whether the caller or the proxy chooses which model to use
Client ID / Client Secret	Credentials for the LLM proxy
Threshold Bytes	Only trims responses larger than this (default 8000 bytes)
LLM Timeout	How long to wait for the proxy before giving up (default 10s)

If anything goes wrong with the LLM call, the response is still delivered — just without the extra trimming.

3. Compress Repeated Structures (off by default — new)

Rewrites repetitive JSON (like arrays of similar records) into TOON, a compact format that carries the same data in fewer tokens. It runs after Payload Sanitization and only when it would actually make the response smaller.

TOON Min Response Size — responses smaller than this (default 500 bytes) are skipped.

Known Issues

Smart Response Trimming is not applied when the policy runs after MCP Schema Validation

There is a known issue where Smart Response Trimming fails to run when the MCP Payload Optimization policy is applied after the MCP Schema Validation policy. In this case, Smart Response Trimming is enabled but has no effect, and responses are returned untrimmed. This is a defect, not intended behavior, and a fix is in progress.

Workaround: Reorder the MCP Payload Optimization policy so that it is applied before MCP Schema Validation.

LLM Proxy credential verification fails even with correct credentials

There is a known issue with the LLM Proxy step used to configure Smart Response Trimming. Credential verification currently fails for all credentials, including valid ones — entering a correct client ID and client secret still returns a verification failure. This is a defect, not intended behavior, and a fix is in progress.

Workaround: None at this time.

Subpolicy cards are not surfaced in the Top Opportunities section

There is a known issue where the new subpolicy cards are not being recommended in the Top Opportunities section. This is a defect, not intended behavior, and a fix is in progress.

Workaround: Open an instance’s modal to view savings directly.