Every tool you expose to an AI assistant costs tokens. The tool's name, description, and parameter schema are included in every request. With 90+ tools, that overhead adds up fast.

CKB's preset system solves this. Here's the math.

The Hidden Cost of Tools

When you connect an MCP server to your AI assistant, the tool schemas are sent with every message. For a typical tool:

{
  "name": "findReferences",
  "description": "Find all references to a symbol with completeness information",
  "parameters": {
    "symbolId": { "type": "string", "description": "The stable symbol ID" },
    "limit": { "type": "number", "default": 100 },
    "scope": { "type": "string", "description": "Optional module scope" }
  }
}

This single tool definition is ~150 tokens. Multiply by CKB's 110 tools and the full toolset is ~17,000 tokens before your AI writes a single line of code.

At current API pricing, that's real money—especially for teams running hundreds of sessions daily.

Preset Token Costs

CKB organizes tools into presets. Here's what each costs:

Preset	Tools	Tokens	Monthly Cost (1000 sessions)*
core	25	~4,100	~$125
docs	34	~5,400	~$160
ops	43	~5,700	~$170
federation	46	~6,700	~$200
refactor	42	~7,500	~$225
review	42	~8,200	~$245
full	110	~17,000	~$510

*Measured on v9.3.1 by serialising the real tools/list response per preset; tokens are bytes/4. Cost assumes $3 per million input tokens and 10 messages per session, uncached.

The core preset gives you 76% fewer tool-schema tokens than full while covering the vast majority of use cases.

What's in Each Preset

Core (25 tools) - The Essentials

The default preset includes:

Navigation: searchSymbols, getSymbol, findReferences, getCallGraph
Understanding: explainSymbol, explainFile, listKeyConcepts
Impact: prepareChange, getHotspots
Compound: explore, understand, batchGet, batchSearch
Utility: expandToolset

This handles 90% of coding sessions. Symbol lookup, reference finding, impact analysis—it's all there.

Review (27 tools) - Code Review Sessions

Adds to core:

getDiff, getFileDiff
getOwnership, getReviewers
getOwnershipDrift

Perfect for PR reviews where you need to know who owns code and who should review changes.

Refactor (33 tools) - Code Cleanup

Adds to core:

analyzeCoupling
findDeadCode
getObservedUsage
detectDuplication

Ideal for refactoring sessions where you're hunting dead code and analyzing dependencies.

Federation (35 tools) - Multi-Repo

Adds to core:

Cross-repo search and navigation
Dependency graph across repositories
Federated impact analysis

For enterprises with multiple interconnected repositories.

Dynamic Expansion

You don't have to choose upfront. Start with core and expand mid-session:

User: "I need to check who owns the auth module"

AI: [calls expandToolset with "review" preset]
    [calls getOwnership for auth module]

"The auth module is primarily owned by @alice (67% of commits)
and @bob (23%). Last significant change was 12 days ago."

The expandToolset tool is included in every preset. Your AI can request additional capabilities when needed.

Measuring Your Savings

Here's a real comparison from a 30-minute coding session:

Without presets (full toolset):

45 messages exchanged
17,000 tokens × 45 = 765,000 tokens in tool overhead
Plus ~200,000 tokens in actual conversation
Total: ~965,000 tokens

With core preset:

Same 45 messages
4,100 tokens × 45 = 184,500 tokens in tool overhead
Same ~200,000 tokens in conversation
Total: ~385,000 tokens

That's a 60% reduction in total token usage for the session.

Where this stops mattering

Prompt caching changes the picture. If your client caches the tool schemas, you pay full price on the first request of a session and a fraction of it afterwards, so the gap above narrows considerably. Presets still help — the first request in every session pays in full, cache writes cost more than reads, and caches expire — but if your setup caches aggressively, treat the numbers above as the ceiling on what presets can save you, not the expected saving.

Choosing the Right Preset

Use core when:

General coding and debugging
Understanding unfamiliar code
Impact analysis before changes
Most everyday development

Use review when:

Reviewing pull requests
Checking code ownership
Identifying reviewers

Use refactor when:

Cleaning up dead code
Analyzing coupling
Planning large refactors

Use federation when:

Working across multiple repos
Tracing cross-repo dependencies
Enterprise monorepo setups

Use full when:

You need everything
Exploring CKB's capabilities
Specialized workflows (security, ops, docs)

Configuration

Set your default preset in the MCP config:

{
  "mcpServers": {
    "ckb": {
      "command": "npx",
      "args": ["@tastehub/ckb", "mcp", "--preset=core"]
    }
  }
}

Or per-project in .ckb/config.json:

{
  "mcp": {
    "defaultPreset": "review"
  }
}

The Bottom Line

Tool presets aren't just about cost—they're about focus. A smaller toolset means your AI spends less time parsing irrelevant options and more time using the right tools effectively.

Start with core. Expand when you need to. Your token budget will thank you.

Links: