I analyzed the token usage of all 51 Palantir MCP tools and found they consume ~37,555 tokens when fully loaded into Claude Code’s context window.
The Solution: Deferred Tool Loading
Instead of loading all tools upfront, use ToolSearch to load only what you need:
ToolSearch query: “select:mcp__palantir-mcp__search_foundry_ontology”
Token Savings by task
| Task | With Deferred | Without | Savings |
|---|---|---|---|
| Documentation lookup | ~150 | 37,555 | 99% |
| Ontology search | 2,800 | 37,555 | 93% |
| Build datasets | ~1,200 | 37,555 | 97% |
Top Token-Heavy Tools
| Tool | Tokens |
|---|---|
| aggregate_ontology_objects | 6,300 |
| create_foundry_rest_api_data_source_webhook | 3,900 |
| search_foundry_ontology | 2,800 |
| create_or_update_foundry_link_type | 2,300 |
The lightweight documentation tools (142-294 tokens) are great for quick lookups without the overhead.