Your agent starts every session blind.

A Claude Code skill (MIT, 76.5k stars) that reads whatever is in a directory — code via tree-sitter AST parsing, plus markdown, PDFs, screenshots, and diagrams via Claude's vision — and builds a persistent knowledge graph you query instead of re-reading files. pip install graphifyy, then /graphify in Claude Code. Everything runs locally with no paid API beyond the Claude you're already paying for; SHA256 caching reprocesses only changed files, and every edge is tagged as extracted, inferred, or ambiguous, so you can see which relationships are facts and which are the model's guess. Headline claim: 71.5x fewer tokens per query than reading raw files, measured on a 52-file mixed corpus.

Full disclosure on the magnitude: created April 3, 76,579 stars in three months, and it trended into our pool last Saturday before re-trending today — this is an established rocket having another moment, not a launch. Reach for it when the context your agent needs is spread across code, schema dumps, papers, and screenshots that no single read can hold. Delete the ritual of re-explaining your project structure at the top of every session. Tradeoff: it's Claude Code-only, and the token math only pays off on larger corpora — on a six-file project the graph costs more than it saves.

A local MCP server (Apache-2.0) that parses your source into a deterministic structural graph — modules, types, routes, storage, dependencies — and exposes it as tools any MCP agent can call: traverse walks reachability, find_path finds the dependency chain between two points, impact_analysis computes the blast radius of a change. The design principle is the whole product: same commit, same graph, no language-model summarization anywhere in the loop. Covers Go, TypeScript, Python, Java, Kotlin, Swift, Ruby, C/C++, PHP, Vue, Svelte, and OpenAPI specs, with framework awareness for Next.js, FastAPI, Django, Spring, and Rails. One curl install, then claude mcp add enola enola.

If graphify is the map of everything with the model's inferences tagged as such, enola is the surveyor's plat: smaller territory, zero guessing. It came out of the builders' own multi-repo project, where the agent could see whichever repo was open and nothing else. Reach for it when your agent keeps confidently editing one service while breaking the contract three repos away. Delete the ARCHITECTURE.md that was stale the week after you wrote it. Tradeoffs: 56 stars and a young project, and the graph only covers what its parsers extract — it's a first step by its own README's admission, not a whole-codebase oracle.

A local CLI (Apache-2.0, 334 stars) that indexes the coding-agent transcripts already sitting on your machine — Claude Code, Codex, Cursor, Pi, OpenCode, Gemini CLI, Factory Droid, Copilot CLI — into SQLite, then gives the agent ctx search and ctx show to retrieve ranked, cited matches with session and event IDs. The pitch names the gap precisely: agents can inspect the current repo, but the discussions, decisions, failed attempts, and test results from earlier sessions are gone every morning. Fully local — no cloud service, no model APIs, no API keys, nothing written into your repos. One curl install plus npx skills add ctxrs/ctx to teach your agent to use it; claims 50x better token efficiency than grepping raw transcripts.

This is the time dimension of today's map theme: enola tells the agent what the code is, ctx tells it what already happened to the code. Reach for it the third time an agent re-attempts the same failed migration strategy you watched it abandon in May. Delete the NOTES.md of past-session context you paste in by hand, and most of the "let me first explore the codebase" preamble spend. Tradeoffs: created in February and trending now rather than brand new, no version number published, and its usefulness is exactly proportional to how much transcript history you've accumulated — day one, it knows nothing.

A CLI (AGPL-3.0, v0.2.0, uv tool install slopo) that catches the duplication traditional tools can't: not copy-paste, but code written similarly, sitting far apart — the same validation logic re-implemented in three modules with different variable names. It embeds every code unit across eight languages, clusters the near-duplicates, and emits a report; the suggested workflow is handing the clusters to your coding agent to consolidate. Indexing is incremental, and an ignore file manages the clusters you've decided to live with.

The README is honest enough not to blame AI — near-duplication predates agents — but the failure mode it targets is precisely what a code generator with no map produces: something that can't see the existing helper writes it again, slightly differently, forever. That makes slopo the audit for the blindness the rest of today's picks prevent. Reach for it quarterly, or before a refactor, on any codebase that grew faster than its review process. Delete nothing yet — run it once and see how bad the sprawl is. Tradeoffs: embeddings come from an external API via LiteLLM (it recommends Voyage's code models), so your source leaves the machine and costs per run, and it's AGPL — fine for internal use, worth a policy check before it touches products you distribute.

A local CLI (MIT) that checks package versions against the OSV and GitHub Advisory databases in parallel, across npm, PyPI, crates.io, Go modules, RubyGems, Maven, NuGet, and seven more ecosystems — blocking criticals, flagging mediums, passing the rest. No hosted service; it calls the public advisory APIs directly. What earns it a slot today is the integration surface: it ships as an MCP server (check_package, suggest_safe_version, compare_versions), as a skill, and as a PreToolUse hook that blocks an unsafe npm install in Claude Code or Codex before the command runs. The maintainer's stated motivation is the theme in one line: it was born out of frustration with AI agents constantly reaching for old versions.

An agent's model of the package ecosystem froze at its training cutoff; every advisory since is invisible until a tool checks. Reach for it if agents install dependencies in your repos with any regularity — the hook mode is the point, because it works even when the agent doesn't think to ask. Delete the manual npm audit step the agent skips. Tradeoffs: ten days old, 19 stars, so this is a fresh bet, not a proven dependency — the case for trying it anyway is that it's MIT, local, and thin enough to read the source in an afternoon.

Off the thread but worth knowing: Apple shipped the Safari MCP server in Safari Technology Preview 247 — a first-party MCP that lets any agent inspect the DOM, read console and network activity, evaluate JavaScript, click, type, screenshot, and run accessibility and performance checks against a real Safari window (claude mcp add safari-mcp-stp -- ".../safaridriver" --mcp, macOS only). Off today's codebase-map thread, but the second browser vendor to ship first-party agent hands after Chrome's DevTools MCP — WebKit debugging without leaving the terminal is now a supported workflow, not a hack. And claude-real-video (github.com/HUANGCHIHHUNGLeo/claude-real-video, MIT, 413 stars, created Monday) gives any LLM working video comprehension by extracting scene-aware deduplicated frames plus transcript from a URL or local file, entirely locally — three days old, so treat it as the fresh bet it is.