DataGrout
Your agent just fetched
10,000 records. Now what?
Three approaches to handling large API payloads in agent workflows. One costs 1,000x more than the other. The formula is visible.
Raw Context
Stuff the entire API response into the agent's context window.
// Agent receives the full payload { "results": [ {"Id": "00Q000000001", "Company": "Acme Corp", ...}, {"Id": "00Q000000002", "Company": "Globex Inc", ...}, ... 9,998 more records ... // ~7.5 MB of JSON → ~1.9M tokens ] }
Roll Your Own
Agent writes filter code across multiple turns. Or build custom hooks/scripts.
// Turn 1: Agent inspects a sample // Turn 2: Writes Python filter leads = [l for l in data if l["State"] == "California" and parse(l["CreatedDate"]) > cutoff] leads.sort(key=lambda x: x["LeadScore"], reverse=True) result = leads[:5] // Turn 3: Verify and format output
DataGrout
Deterministic tools extract exactly what the agent needs. Zero LLM credits.
// Single MCP tool call chain { "name": "data-grout@1/frame.filter@1", "arguments": { "cache_ref": "rc_leads_abc123", "where": {"State": "California"} } } // → frame.sort → frame.slice → frame.select // Result: 5 records, 6 fields = ~340 tokens
Cost Calculator
Adjust the parameters to see how costs scale with your data.
Empirical Benchmark Results
Measured on generated datasets using DataGrout's deterministic tool suite. Model: Claude Opus 4 at $15/1M input tokens.
Experiment 1: Salesforce 10k Leads
| Approach | Input Tokens | Cost | Turns | Accurate | Savings |
|---|---|---|---|---|---|
| Raw Context | 1,880,425 | $28.22 | 1 | No | baseline |
| Roll Your Own | 3,320 | $0.1548 | 3 | Yes | 99.8% |
| DataGrout | 348 | $0.0265 | 1 | Yes | 100.0% |
Experiment 2: Order Trend Analysis (5k orders)
| Approach | Input Tokens | Cost | Turns | Accurate | Savings |
|---|---|---|---|---|---|
| Raw Context | 519,475 | $7.83 | 1 | No | baseline |
| Roll Your Own | 4,300 | $0.2370 | 3 | Yes | 99.2% |
| DataGrout | 401 | $0.0585 | 1 | Yes | 99.9% |
Experiment 3: Multi-Source Integration (SAP + Salesforce)
| Approach | Input Tokens | Cost | Turns | Accurate | Savings |
|---|---|---|---|---|---|
| Raw Context | 350,150 | $5.30 | 1 | No | baseline |
| Roll Your Own | 6,150 | $0.3248 | 4 | Yes | 98.2% |
| DataGrout | 750 | $0.0738 | 1 | Yes | 99.8% |
The Pattern
Deterministic first. LLM second. This isn't new — a recent analysis showed that awk scripts alone can collapse 108,894 bytes of terminal noise to 37. But terminal noise is the easy case.
The hard case is structured data — API responses, database results, integration payloads. You can't awk your way through 10,000 JSON records. That's where DataGrout's tool suites come in.
The Tools That Make This Work
DataGrout's deterministic suites handle the heavy lifting before any LLM token is spent.
frame.* Suite
Filter, sort, group, pivot, slice, join — columnar ops on record sets. Accepts cache_ref so large payloads stay server-side.
0 credits · deterministicdata.* Suite
Pick, omit, flatten, aggregate, unique — pure JSON transformations. Shrink payloads to just the fields you need.
0 credits · deterministicprism.refract
Natural-language data reshaping. "Group by customer, sum totals" — generates and caches reusable Starlark transforms.
LLM-backed · skill cachingprism.chart
Turn raw numbers into a sparkline + text summary. The agent gets a 50-token visual instead of 500k raw values.
2 credits · visual compressionmath.describe
Statistical summary: mean, median, percentiles, histogram. Deterministic analysis of numeric columns.
0 credits · deterministicprism.paginate + cache_ref
Keep large results server-side entirely. Agent pages through results without ever loading the full set into context.
0 credits · server-sidePricing
5,000 free credits/month. Deterministic tools (frame.*, data.*, math.*) cost 0 credits — only the gateway base fee applies. LLM-backed tools charge 1 credit + passthrough model cost at 1.2× margin. See full pricing →