I spent part of this week poking through the trending dev tools on GitHub, and a pattern jumped out: almost everything new and interesting isn’t a tool for you — it’s a tool for your coding agent. Memory systems, search backends, sandboxed tool output, pre-packaged workflows. The new developer experience (DX) is agentic developer experience (ADX?). Here are the standouts, plus a few solid tools that still respect the human in the loop.

Give your agent a memory: claude-mem

The biggest pain point with coding agents is still amnesia — every new session starts from zero. claude-mem (Apache-2.0, 100k+ stars on GitHub) attacks that head-on: lifecycle hooks capture what your agent does each session, compress it with AI, and inject the relevant bits back into future sessions.

It works with Claude Code, OpenClaw, Codex, Gemini, Copilot, OpenCode and more, stores everything locally in SQLite plus a Chroma vector database, and exposes memory through MCP search tools with a three-layer workflow — search an index first, pull timelines around interesting hits, and only then fetch full details. Per its README that progressive disclosure saves roughly 10x the tokens versus dumping everything into context. It’s a hosted-memory service with a free tier by default (it asks you to sign in), but you can point it at your own keys or run it fully local. If you live in an agent session all day, this is the one to try first.

Stop burning tokens on grep: XERJ

Agents that search with grep pull whole files into context — the XERJ project measured up to 1M input tokens on one codebase that way. XERJ (Apache-2.0, created June 2026) is a local AI search engine you install with one command, point at a folder with xerj autoindex, and wire into Claude Code or Cursor over MCP. The agent queries it and gets the exact passage with file and line number instead of the whole file.

The project’s own case study — real claude -p token counts across 8 tasks in 4 languages — claims 2.7x fewer output tokens than a grep-driven agent at the same 16/16 solve rate, and 26x fewer than an agent working from memory alone. Treat those numbers the way you’d treat any vendor benchmark (their README does publish the methodology and the honest limits), but the underlying idea is sound: an Elasticsearch-compatible API on a local index that your agent can query is a lot cheaper than reading the repo by hand. xerj init sets up the MCP wiring and a skill for you.

Shrink what your tools say: Context Mode

Even with good search, tool output is a token firehose — one agent roundup this week surfaced Context Mode, an MCP server that runs tool output through a sandbox and hands the agent a digest instead of the raw dump. The example making the rounds: 315 KB of output compressed to 5.4 KB — a 98% reduction. I haven’t kicked the tires on this one myself (it’s per the trending roundup that surfaced it, so file the numbers under “their demo”), but the direction is right: context hygiene as infrastructure, not something each agent improvises.

Skill packs: stop re-teaching the same workflows

Two related projects package up how agents should work:

• mattpocock/skills (MIT) — a collection of coding-agent skills that walk an agent through a disciplined loop: rephrase the task, write a spec, break it into tickets, then TDD, then review. Basically, the workflow you’d enforce on a junior dev, pre-packaged for an agent.
• ECC — goes maximalist with 71 agents and 311 skills running a plan → test → implement → review → verify → remember → improve cycle.

I’m skeptical of the “hundreds of skills” approach — skill bloat is its own tax — but the small, opinionated pack is worth a look.

Reverse engineering, agent-style: REA

The most dramatic star surge of the week was REA (Reverse Engineer Anything), which one GitHub ranking had gaining roughly 248k stars in a week. It’s an MCP server that lets Claude Code, Codex, or Cursor drive reverse-engineering tools like Hopper, Ghidra, or IDA to explain how an application works — without the source code. Trend-chasing repos spike and crash all the time, so judge the tool, not the star count. But the idea of an agent that can poke at a binary and explain itself has obvious appeal for anyone who’s ever inherited a mystery executable.

And a few for the humans

Not everything this week was agentware:

• SvelteKit 3 (free, MIT) — the new major version moves configuration into vite.config.ts. If you build with Svelte, worth a read before your next upgrade.
• Cloudflare’s cf CLI (free, Apache-2.0, open beta) — one command-line tool covering the public Cloudflare API and Workers, instead of juggling wrangler plus hand-rolled API calls. Very appealing for the Cloudflare-heavy homelab crowd.
• WSL containers (wslc) (free, MIT) — build, run, and deploy Linux containers on Windows through WSL. Ships with WSL itself.
• Node.js 26.11 — the new Current-line release adds a --process-timeout flag for diagnosing stuck processes, better perf_hooks histogram snapshots and diffs, configurable HTTP/2 connection window sizes, and updates to the built-in SQLite API naming. 26.11.1 followed quickly to revert a few doc/build changes, so the runtime features all stand.
• AG-UI 1.0 (free, MIT) — an open protocol that connects AI agents to application front ends. Still early, but the agent-to-app layer is one of the genuinely open problems right now, and a shared protocol beats fifty bespoke bridges.

The takeaway

The theme of the week is unmistakable: we’re building a second toolchain, parallel to the human one, whose users are agents. Memory, search, output hygiene, workflow skills, and agent-to-app protocols are the new DX basics. If you run coding agents at all, claude-mem and XERJ are the two I’d actually install this weekend. Everything else is worth a star and a revisit in a month.