← Other Products

The capability diet: tool pruning for agents

Agents running large MCP/tool stacks load hundreds of tool descriptions they never call. This service compiles cross-session usage telemetry offline into per-task-type load manifests: the tools a task type actually uses stay loaded, the rest are demoted — reversibly, in-session, via a JIT fetch (diet_fetch_tool) the moment one is needed. It includes exposure accounting (description tokens loaded vs the full catalog) and poisoning-resistant compilation: distinct-session scoring, per-agent contribution capped at 25% of a task type's sessions, and a promotion floor of at least 2 distinct agents. Manifests are integrity-signed and versioned. Lab proof (28-tool stack): 11 kept / 17 demoted per task type, -50.9% description tokens/session at equal task success (360/360), 48/48 JIT recoveries.

This product is built and being readied for launch — checkout opens when it clears its final gates.

$9 per fleet / month.

For agents

API endpoint (JSON): https://textforagents.com/products/capability-diet-tool-pruning/ · OpenAPI spec · Agent guide (llms.txt)

Support: support@textforagents.com