Coding agents waste 10x tokens reading files: 955-job benchmark of a code-graph MCP server
Classeve · reddit · 2026-10-01
The Context Zero Engine team open-sourced an MCP server that indexes repos into a code graph (symbols, calls, effects, contracts, blast radius) so agents query structure instead of reading files. Their reproducible benchmark across 18 pinned public repos, 11 languages and 955 jobs shows median per-task token use dropping from 13,321 to 2,221 (83% less), pooled usage from 23.3M to 2.3M tokens (10.1x), and 6.7x savings even vs an oracle baseline. On 101 jobs the graph cost more because it returns callers, effects and contracts. Ships 61 MCP tools across 13 analysis engines, runs locally on PostgreSQL, works with Claude Code.
More from coding & agent
- Developer wakes up to 6.5 hours of agent-completed work — calls it 2 weeks of his time — therealdanvega · 2026-10-01
- Ex-OpenAI Researcher Launches Jev: A System One Model for Structured Decisions, 100x Faster — KhuyenTran16 · 2026-10-01
- Jev vs PydanticAI: control the decision before the model responds, with confidence scores — KhuyenTran16 · 2026-10-01
- Dev builds MCP server to fetch podcast transcripts as a high-quality info source — vista8 · 2026-10-01
- Corbenic AI launches Galahad beta: disk-persistent KV cache reuse across requests and restarts — MindPsychological140 · 2026-10-01
- How Abridge turned clinician feedback into a measurable agent eval system — LangChain · 2026-10-01