§ proof · context efficiency
Nukez Context Management
Token cost per operation.
These are the compact comparison numbers used across the docs. The comparison is most useful when each surface is measured against the same workload and refreshed as new trace runs are collected.
NukezAgent
The parent agent sends one compact delegation request. File bytes, tool history, signatures, and receipts stay inside the hosted storage specialist.
PyNukez SDK
Python-native lifecycle calls keep protocol work out of prose, but the full SDK tool surface is still available to the calling agent.
MCP
A lean hosted tool surface for MCP-speaking clients. Current frontend baseline is shown; fresh live trace evidence can replace it as new runs are collected.
HTTP Gateway
Maximum portability, maximum context exposure. Agents must carry discovery docs, request shapes, envelope details, and raw JSON responses.
§ what the traces prove
The shape matters more than a single number.
Raw traces count the whole autonomous loop: prompt, messages, tool schemas, tool results, fetched documentation, and final response. The public tok/op number is the simpler product comparison. Both are useful, but they should not be collapsed into one claim.
- Best parent-context shape
- NukezAgent
- Best language-native path
- PyNukez SDK
- Most portable path
- HTTP Gateway
- Needs fresh saved JSON
- MCP
- Parent payload
- 0 file bytes
- Live sample
- 3,422-3,485 input tokens
- 13-call run
- 0-token context growth
- Latest success
- 9 LLM calls
- Peak input
- 10,370 tokens
- Total input
- 80,607 tokens
- Tool surface
- 15 hosted tools
- Audit runbook
- static + session modes
- Fresh trace
- replace baseline
- Full success
- 21 LLM calls
- Peak input
- 37,864 tokens
- Total input
- 586,749 tokens
§ located evidence
Trace artifacts and status.
This table is intentionally provenance-first. If a surface has a saved JSON trace, it says so. If it only has the audit harness, handoff notes, or a pending live trace, it says that too.
| Surface | Source | Result | Steps | Total input | Peak input |
|---|---|---|---|---|---|
| NukezAgent | nukez-traces timeline, 2026-03-25 | 13 calls, constant parent window | 13 | 44,785 est. | 3,485 |
| PyNukez SDK | sdk_traces_anthropic_20260411_114022.json | task_complete | 9 | 80,607 | 10,370 |
| HTTP Gateway | gateway_traces_anthropic_20260327_174345.json | task_complete | 21 | 586,749 | 37,864 |
| HTTP Gateway | gateway_traces_anthropic_20260411_153613.json | stress / non-canonical | 40 | 1,422,852 | 53,200 |
| MCP | context_audit.py + WALKTHROUGH.md | harness available | 25 simulated | pending | pending |
Live trace handoff records 3,480 -> 3,480 growth and avg 3,445 input/call.
Latest located SDK success trace: request, transfer, confirm, provision, upload, download, list.
Full raw-gateway autonomous store+verify path. Exceeds 32K once the agent carries docs and raw responses.
§ comparison math
8.5x lower
162 vs 1,380 tok/op
2.7x lower
510 vs 1,380 tok/op
5.6x lower
162 vs 904 tok/op