Token costs rise fastest at the point where enterprise AI moves from pilot to production. The model is billed for every tool definition loaded into a request and every row a query hands back, and those charges repeat on each turn of a conversation. The most effective way to lower that bill is to narrow what the model sees before it runs, using a few purpose-built tools in place of a full catalog of generic ones.
This guide covers seven ways custom MCP tools cut those costs, with measured numbers from an internal CData benchmark of 56 live runs.
CData Connect AI and custom MCP tools overview
The Model Context Protocol (MCP) is an open standard that connects AI agents and assistants to external systems through one interface, so a single integration replaces a custom build for every pairing. A custom MCP tool is an admin-defined tool on that interface: a named, parameterized query that returns one governed shape of data.
Three things drive the bill on a governed enterprise query:
Tool definitions. Most MCP clients load every definition into context before the model reads the question, so unused tools bill on each turn.
Round trips. Discovery chains and chained queries push results back through context repeatedly. Code execution with MCP took one workflow from 150,000 tokens to 2,000 by keeping that traffic out of the model.
Payload width. A tool that returns 70 columns bills for 70 columns, however many the question needed.
CData Connect AI handles all three in the data layer, before an agent ever sees a prompt, with five architectural levers doing the work:
Pre-joined virtual views: Derived Views turn three data calls into one, so joins and filters across sources resolve before the agent asks.
Scoped data catalogs: Workspaces decide which tables and views an agent can see at all, so a session stays pointed at the sources that matter.
Curated tool bundles: Toolkits group tools into named, workflow-specific sets, and each set publishes as its own MCP endpoint.
Custom scoped tools: Custom Tools replace a 70-column schema with the six fields a question needs, and they take declared parameters so an agent calls them with arguments.
Cached query results: Jobs serve reads from a managed store on a refresh schedule instead of hitting live systems on every call.
An internal benchmark put a price on that. One question spanning Salesforce, Snowflake, and ServiceNow cost about 60 cents per query when the model explored the three sources itself. Routed through a single Custom Tool, the same question cost under 3 cents, a drop of 97.6%.
That result held across 56 runs on Claude Sonnet 4.6. The full results are published, and you can rerun the harness against your own sources using the replication guide and the benchmark repository.
1. Direct tool access to live APIs
A direct MCP connection lets an agent work against the live system, whether that’s Shopify, an ads platform, or a warehouse in your own network. The agent reads and writes through one interface, which removes the clipboard step where somebody exports a report, opens the chat window, and pastes it in.
Those two paths price out very differently. A pasted export bills every column of every row at input rates, whether the question needs 12 rows or 12,000. A custom tool inverts that. The agent sends parameters, Connect AI runs the query against the live source, and only the matching rows come back.
Before: export a full month of order rows, paste them into the prompt, then ask which enterprise accounts are churning.
After: call one tool with a date range and a customer segment, and receive 50 rows across six columns.
The overhead in the first path is paid on every rerun, and it grows with the size of the export rather than the size of the question.
The internal benchmark measured both paths on the same question:
Path | Tokens | Tool calls | Cost per query |
Raw discovery across three sources | 183,541 | 22 | $0.596 |
One Custom Tool | 4,427 | 1 | $0.027 |
Speed moved with cost, from 242.8 seconds down to 19 seconds per query. At 100,000 queries a month, the gap is roughly $57,000 in token spend avoided.
Live access also drops the staleness problem. An export starts ageing the moment it’s created, and an answer drawn from stale rows bills twice: once for the wrong run, once for the rerun.
2. Structured web content ingestion
Web and API sources return payloads built for browsers and developers rather than for models. Raw HTML carries navigation, markup, and scripts. Raw JSON carries envelopes, nested objects, and null-heavy fields. All of it bills at input rates.
Structuring that content before it reaches the model means preprocessing the response so only business-relevant values transmit. In practice, structured output looks like:
Typed rows and columns in place of nested JSON
Named metrics in place of full record dumps
Cleaned text values in place of markup-wrapped fragments
Consistent keys across sources that use different native formats
Connect AI models every connected source, including SaaS and API sources, as tables and views in a single catalog. A query against an API source returns the same tabular result as a query against a database, so the model reads values instead of parsing structure. That shortens the response and removes a class of interpretation errors.
3. Business-context filtering for relevant metrics
Business-context filtering means transmitting only the metrics and attributes a specific question needs. A default Salesforce Account tool exposes more than 70 fields. Most operational questions read a handful, and the unused fields bill on every call.
The benchmark’s Custom Tool replaced that 73-column schema with a six-column definition covering the account, industry, revenue, and opportunity values the query needed. The tool definition shrank by roughly 80%, and the result rows shrank with it.
Filtering at the tool layer carries a governance benefit at no extra cost. Columns absent from the tool definition never reach the model, so sensitive attributes stay out of context by design rather than by prompt instruction.
4. Retrieval-augmented generation with selective snippets
Retrieval-augmented generation (RAG) is an architecture where the model queries an external vector database and receives only the top-matching passages for the question at hand, instead of the full corpus. The token argument for RAG and the token argument for custom tools are the same argument applied to different data shapes.
Documents suit vector retrieval. Systems of record suit query retrieval, where the filter belongs in a WHERE clause and the limit belongs in a TOP or LIMIT. A custom tool encodes both, so an agent asking about at-risk accounts receives 50 scored rows rather than a table scan.
Teams running both patterns get the cleanest results when each store answers what it’s shaped for. Vector search handles contracts, tickets, and policy text. Custom tools handle the CRM, the warehouse, and the ledger, with joins resolved before retrieval rather than reassembled by the model.
5. Multi-integration hubs to consolidate endpoints
Connecting four MCP servers to one agent loads four tool registries, four authentication surfaces, and four sets of schema metadata. That overhead repeats on every turn, which is why distributed tool architectures produce token bills that scale with connections rather than with questions.
A single governed endpoint collapses the repetition. In Connect AI, Workspaces publish a scoped catalog through one MCP endpoint, and Derived Views resolve cross-source joins in SQL before the model sees anything. The internal benchmark measured Workspaces at $0.049 per query and Derived Views at $0.146, against $0.596 for the raw path.
Consolidation also improves tool selection. Anthropic’s testing on deferred tool loading recorded an 85% cut in token usage and accuracy on MCP evaluations rising from 49% to 74% on Opus 4, once tool definitions stopped competing for context. Fewer wrong answers means fewer paid retries.
6. Pre-filtering and anomaly-driven alerts
Pre-filtering processes data at the integration layer and sends the model exceptions and summaries in place of full streams. A monitoring agent reading 10,000 raw events pays for 10,000 raw events. The same agent reading a ranked list of 40 threshold breaches pays for 40 rows.
Custom Tools carry the aggregation in their SQL, so counts, deltas, and threshold logic run server-side. Jobs are CData’s scheduled cache tasks, copying selected tables to a managed store so recurring queries skip the live round trip. The internal benchmark measured that cached path at $0.075 per query, an 89.2% reduction.
Narrow payloads sharpen the agent’s reasoning as well as the invoice. Anthropic’s research on context engineering describes a finite attention budget, where recall accuracy declines as context fills across every model tested.
7. Governance and environment separation controls
Governance controls limit what a session can reach, which caps both risk and spend. An agent that can query any table can also scan any table, and an unbounded scan is billed like any other request.
A workable control set for MCP deployments:
Control | What it limits |
Role-based access | Which users and agents see which sources |
Scoped endpoints | Which assets a given agent can call at all |
Read-only staging | Whether a session can modify source systems |
Column and row policies | Which fields and records enter context |
Audit logging | What was queried, by whom, and at what cost |
Connect AI applies these at the platform rather than per client. Access runs through OAuth and single sign-on with personal access tokens for programmatic callers, Workspaces separate development assets from production ones, and every query is logged.
Separating environments matters for cost as much as for compliance, since exploratory sessions and scheduled production runs have different query shapes and belong on different tool surfaces.
Frequently asked questions
What are common methods to reduce LLM token usage with MCP tools?
Shorten the tool menu, scope tool definitions to the columns a workflow reads, pre-join across sources so the model makes one call instead of several, cache recurring queries, and set per-session limits on what an agent can reach.
How does retrieval-augmented context improve token efficiency?
Retrieval sends the model only the passages or rows that match the question, so a prompt carries a scoped result rather than a full corpus or table. The saving grows with the size of the underlying store.
Can governance controls prevent unexpected token overconsumption?
Yes. Scoped endpoints, role-based access, and read-only surfaces bound the volume any single session can pull, and audit logs make cost spikes traceable to the query that caused them.
What is the best order to implement MCP token-saving strategies?
Measure a baseline first, then scope the catalog each agent sees, trim tool definitions on high-use sources, pre-join recurring multi-source queries, cache anything that tolerates hourly freshness, and apply governance limits last.
How do compact response formats affect model performance?
Compact formats hold answer quality when the fields a workflow needs stay present, and narrower context often reads more accurately. In Anthropic’s testing, orchestrating tool calls in code cut average usage from 43,588 tokens to 27,297 on complex research tasks while reducing errors.
Spend less on every agent action with CData Connect AI
Every extra column and every unused tool an agent can see turns up on the invoice. Narrow both, and the same business question costs cents instead of dollars. CData Connect AI gives you that control across hundreds of governed sources through one managed MCP endpoint.
Start a free trial and reduce your token costs today.
Your enterprise data, finally AI-ready.
Connect AI gives your AI assistants and agents live, governed access to 350+ enterprise systems so they can reason over your actual business data, not just what they were trained on.
