Home / Blog
Blog
Context engineering for AI coding agents
Benchmarks, guides and release notes on giving Claude Code, Cursor, Codex and other MCP agents the right code — with the numbers and the caveats.
What is a context engine?
A context engine decides which code reaches your AI agent. How indexing, retrieval and snippet assembly cut tokens and improve answers — and how it differs from RAG and grep.
Release notesFinder 1.2.0: 20 ms identifier lookups, one-file Finder Kit, finder update
What changed in Finder 1.2.0 and what we measured: local identifier lookups (1.2–1.7 s → ~20 ms), compact FCP/2 answers (6.6k → 1.2k chars), BM25 back on for every kit request, a Finder Kit that installs as one skill, a guard against indexing whole drives, and a one-word finder update.
BenchmarksFinder 1.0.0 speed benchmark: 0.5–0.7 s new questions, 35 ms repeats
Finder 1.0.0 against another local context-engine MCP server on a Flutter and a Rust repo: latency, ground-truth recall, output size, memory and CPU — and where Finder did not win.
All posts
Finder 0.0.14 vs GitNexus: code search benchmark (October 2026)
Eight ground-truth questions on a real Rust/TypeScript repo: Finder 0.0.14 vs GitNexus 1.6.9 vs an agent with only grep. Recall 0.93 vs 0.27, tokens, latency and caveats.
Release notesFinder 0.0.11: sharper file-finder, auto-start, lighter disk
What changed in Finder 0.0.11 and what we measured: symbol-unit rerank (IoU 0.24 → 0.74), stdio auto-start for 15 harnesses, per-call index wait, 893 MB less disk — and the one thing that got slower.
BenchmarksNo MCP vs Finder vs GitNexus vs CodeGraph: code retrieval benchmark
Eight ground-truth questions on a real Rust/TypeScript repo: an agent with only grep and file reads, versus Finder, GitNexus and CodeGraph over MCP. Recall, latency and tokens — with the caveats.
GuidesHow to cut token usage in Claude Code, Cursor and Codex
Most tokens an AI coding agent spends go to reading files it didn't need. Five practical ways to shrink context — and how a context engine returns snippets instead of files.
BenchmarksFinder 0.0.10 benchmark: 0.95 recall at ~770 ms
Retrieval benchmark of Finder 0.0.10 against a popular hosted context-engine MCP server: latency, ground-truth recall, precision and file-level hit rates — including where Finder loses.
SetupSet up a code search MCP server in Claude Code, Cursor, Codex and 12 more
Step-by-step: install Finder, connect it as an MCP server to Claude Code, Cursor, Codex, VS Code, Gemini CLI, Windsurf, Kiro, Cline, Zed and others, and approve your first workspace.
ExplainersSemantic code search vs grep for AI coding agents
Grep matches text; semantic code search matches intent. When each wins, why agents need both, and how embeddings plus a call graph beat either alone.
ExplainersImpact analysis for AI agents: know the blast radius before you edit
Why coding agents break things they didn't read, and how call-graph impact analysis gives them callers, dependents and tests to run before an edit — with no LLM call.
ExplainersWhat is an MCP server for code search?
The Model Context Protocol lets AI agents call external tools. Here's what a code-search MCP server does, what tools it should expose, and what to check before you install one.
No posts match.
Put the right code in your agent's context
One command installs Finder and wires it into your harness. Start with 40 free credits.