Tools consume context before the task begins
A model must read enough of each tool contract to decide whether and how to call it.
An MCP server may expose dozens of valuable capabilities. In hosts that preload the catalog, every description and input schema consumes context before the agent sees the request. Similar names can also make routing harder: the model must distinguish several ways to inspect, plan, or review before it can work.
The right response is not to remove advanced capabilities. It is to choose a useful default surface and let specialized work opt into the broader catalog. The core profile should cover the normal loop from setup and context through validation and handoff.
Brief makes the catalog budget testable
The package records maximum tools, schema bytes, catalog bytes, and estimated tokens in a versioned JSON contract.
Brief’s core profile allows up to 24 named tools, 62,000 total catalog bytes, and 15,000 estimated tokens. The full profile allows up to 64 tools, 130,000 bytes, and 31,000 estimated tokens. A separate assertion limits the core estimate to half of the full profile.
The estimator is intentionally simple: it divides description and input-schema bytes by four and rounds up. That is not a tokenizer-accurate usage report. It is a stable engineering guardrail that catches catalog growth during releases.
- Core: no more than 24 tools
- Core: no more than 15,000 estimated catalog tokens
- Full: no more than 64 tools
- Full: no more than 31,000 estimated catalog tokens
- Core estimate: no more than 50% of the full profile
Choose the default around a complete job
A small catalog is useful only if it still covers the workflow users perform most often.
Brief’s core list includes connection status, context, code mapping, test discovery, mission planning, frontend proof, local diff review, artifact publishing and comments, learning, handoff, and receipt verification. Together they cover a complete repository-change and review loop within the smaller profile.
Specialized diagnostics, governance, marketplace, and administrative surfaces remain available in the full profile. This avoids forcing every user to pay the discovery cost of a rare workflow while keeping an explicit route to it.
Treat smaller as a hypothesis, then measure it
Catalog size is an input metric; task completion, routing errors, latency, and review quality are the outcomes.
A lower token estimate does not automatically make an agent better. If the core profile hides a tool needed for a common task, users will lose time or invent a workaround. Teams should track missed-tool requests and tool-selection errors alongside catalog growth.
Brief’s budget is useful because it makes the tradeoff visible and regression-testable. It is original product data with a narrow claim: the default catalog is deliberately bounded. Broader claims about model accuracy require separate evaluations.
Common questions
Questions worth asking next
Does a 15,000-token budget mean every Brief task uses 15,000 tokens?
No. It is a ceiling for the estimated tool catalog contract, not total task usage and not a tokenizer-accurate measurement of a specific host.
Are full-profile tools removed from Brief?
No. The full profile remains available for specialized work; the core profile keeps the everyday surface focused.