Help your AI coding agent understand your repo
without burning every token.
EngramX builds a tiny structural index of your codebase so Claude Code, Cursor, and Cline can answer questions about it without re-reading every file. Local SQLite. No account. No telemetry.
Across an 87-file reference codebase, EngramX cut token cost by 53–89% per file, depending on the file. Two files got worse — the structural index added more overhead than it saved. The benchmark script is in the public repo. Run it against your own repo in five minutes. Don't take our word.
What it does,
in plain language.
Three short paragraphs for three different readers. Pick the one that matches you.
If you've never heard of context tools. When you chat with an AI coding assistant about your project, it has to read your files to answer. Reading the same files over and over costs tokens — which costs money or hits a rate limit. EngramX scans your project once into a small local database, then quietly hands the assistant a compact summary instead of the raw file. The assistant still ships the same edits; it just spends fewer tokens getting there.
If you're comparing EngramX to Cursor's built-in indexer or Aider's repo-map. EngramX is a local SQLite knowledge graph plus a hook layer. It intercepts the agent's Read tool call at the tool boundary and returns a structural packet — symbols, signatures, neighbours, recent changes — rather than the full file body. It runs across Claude Code, Cursor, Cline, Continue.dev and any MCP-compatible host, so one graph serves every IDE you use. It also records mistakes from your git revert history and warns the agent before it re-introduces a bug you already fixed.
If you got here from cirvgreen.com looking for a WordPress plugin. You're probably not the buyer for this — that's fine. EngramX is a developer tool for people running Claude Code or Cursor against a codebase. If that's not you, scroll to the bottom and we'll point you at the right plugin.
Try it in
60 seconds.
One install, one query, one hook. No account, no API key, no telemetry. Apache 2.0.
$ npx engramx init $ engram query "where is X defined?" $ engram install-hook --auto-reindex ↳ .engram/graph.db built · hook armed · ready
No account, no API key, no telemetry. Apache 2.0. The hook writes to your project's local Claude Code settings file (gitignored by default) and you can remove it with engram uninstall-hook any time.
What you can do
with it.
The commands you'll actually use. Full reference is on GitHub.
| Command | What it does |
|---|---|
engram init |
Scan the project and build the local .engram/graph.db structural index. Run once per repo. |
engram query "…" |
Ask a question about the codebase — definitions, callers, neighbours. Returns a compact structural answer. |
engram gen |
Generate an agent-ready context packet for the current task. Use --task bug-fix, --task feature or --task refactor to bias the output. |
engram gods |
Show the highest-connectivity files in the repo — the modules everything else depends on. |
engram mistakes |
List bugs the agent already introduced and reverted in this repo, with the file paths and patterns to avoid. |
engram learn "…" |
Record a fact, decision, or convention you want future agent runs to respect. |
engram install-hook |
Wire EngramX into Claude Code's hook contract so context injection is automatic. Add --auto-reindex for live updates. |
Works inside your
existing tools.
One graph, every IDE. No vendor lock-in.
EngramX ships an MCP server out of the box — run it with npx -y engramx serve /path/to/project and any MCP-compatible host can call its six tools (query_graph, god_nodes, graph_stats, shortest_path, benchmark, list_mistakes). For IDEs without MCP yet, there are direct integrations: a Continue.dev @engram provider, an OpenVSX VS Code extension, and a Cursor MDC bundle.
It remembers
what broke last time.
The mistake-guard hook reads your git revert history so your AI agent stops re-introducing yesterday's bug.
Every time you revert a commit, EngramX quietly records what was reverted, where, and when. The next time the agent is about to write code that matches one of those patterns, a PreToolUse hook fires — EngramX's contribution to Claude Code's hook contract — and the agent sees a warning before the file is written. It's bi-temporal, which means EngramX remembers when the mistake happened and when you fixed it, so warnings fade as the codebase moves on.
Disabled by default until you opt in with engram install-hook. Runs in permissive mode — it warns, it doesn't block. Set ENGRAM_MISTAKE_GUARD=0 in your environment to turn it off entirely. The whole thing is local; no revert history ever leaves your machine.
Apache 2.0. On GitHub.
Built in public.
No license fee. No paid tier. No "community edition." It's the whole thing, free, with the source open.
EngramX is a single-maintainer open-source project published from Dubai. The npm package is engramx; the GitHub repo includes the benchmark script that produced the 53–89% range, so the claim is reproducible against any codebase you choose. CHANGELOG and issues are open. PRs welcome.
Looking for a WordPress
plugin instead?
No problem. EngramX is a developer tool. Here are the actual plugins.
EngramX is for developers using AI coding agents. If you landed here from cirvgreen.com looking for a WordPress plugin, you probably want one of these:
- Cirv Box SchemaValid JSON-LD for posts, products, organization, FAQ and breadcrumbs. Passes Google's Rich Results test on the first try. Details →
- Cirv Guard AccessibilityWCAG 2.1 AA scanner with one-click fixes for alt text, headings, contrast and focus order. Details →
- Cirv Pulse Core Web VitalsLighthouse inside wp-admin. Tracks LCP, INP and CLS per URL and flags regressions early. Details →
- Cirv Comply ConsentCookie banners that survive a GDPR, CCPA, LGPD or PIPEDA audit. Google Consent Mode v2 wired up by default. Details →