The local-first AI coding stack is forming: Recall + Junction in one week

What happened

In the same 72-hour window, two open-source projects targeting the AI coding workflow landed on the front pages of Hacker News and GitHub Trending. Raiyan Yahya's Recall (115 points on HN) ships a local SQLite memory layer for Claude Code, exposing six MCP tools and persisting everything to a `.recall/` directory inside your repo. Junction (518 stars, by Plaer1) is a VS Code chat sidebar that talks to local AI coding agents — Ollama, llama.cpp, or anything that speaks the OpenAI-compatible API on `localhost`.

Neither project is technically novel in isolation. MCP servers have proliferated since Anthropic open-sourced the protocol last year, and there are at least a dozen "chat with your local model in VS Code" extensions. What's notable is the simultaneity: two unrelated devs, two different layers of the same stack, both shipping the local-first version of a piece that the dominant tools (Cursor, Copilot, Claude Code itself) treat as cloud-native by default.

The pitches are almost mirror images. Recall's README leads with "zero cloud, your data stays on your machine." Junction's leads with "no API keys, no telemetry, no accounts." Both projects ship as single-binary or single-extension installs. Both assume you already have the model layer figured out — Recall sits on top of Anthropic's API via Claude Code; Junction sits on top of whatever local inference server you're running.

Why it matters

The AI coding tool market spent 2024 and most of 2025 consolidating around three cloud-first incumbents: Cursor, GitHub Copilot, and Claude Code. Each one bets that the right architecture is your editor talking to their backend, with context, history, and tool-routing handled server-side. The economic logic is straightforward — every byte of context they manage is a byte they can charge for, and every memory store they own is a moat against you switching.

Recall and Junction are evidence that a non-trivial slice of senior developers is willing to give up the convenience of managed context in exchange for owning the substrate. It's the same pattern that produced self-hosted Sentry alternatives, the Mastodon/ActivityPub wave, and the slow migration from Notion to Obsidian — once a category matures, a faction always defects to the local-first version.

The technical case isn't symmetric across all three layers, though. Model weights are the hardest piece to localize — Qwen 2.5 Coder 32B and DeepSeek Coder V2 are competitive with GPT-4o on small-to-medium tasks, but still trail Claude 3.5 Sonnet and GPT-5-class models on long-horizon agentic work. Memory and UI, by contrast, are trivially localizable. There's no reason a chat sidebar needs to phone home, and no reason your project's accumulated context needs to live in someone else's database. Recall and Junction are picking the two layers where the local-first argument is strongest and the model-quality gap is irrelevant.

The community reaction tracks this. The top HN comment on Recall asks the obvious question — "why isn't this just built into Claude Code?" — and the answer in the thread is roughly: because Anthropic's commercial incentive is to make context portable across *their* surfaces, not yours. Junction's GitHub issues are full of requests for Ollama model auto-detection and offline-only mode, which tells you who's actually installing it.

What this means for your stack

If you're running Claude Code daily, Recall is worth installing this week. The six MCP tools (`remember`, `recall`, `forget`, `search_memories`, `list_memories`, `summarize`) cover the actual use cases — persistent project context, decision logs, "why did we do it this way" notes — and the SQLite store means you can grep it, back it up, or sync it across machines with whatever tooling you already use. The integration cost is one MCP server entry in your config; the lock-in cost is zero because the schema is human-readable.

Junction is a harder sell if you're already getting good results from Cursor or Copilot. The local-model coding experience is still rough around the edges — multi-file edits, repo-wide refactors, and long-running agentic tasks all degrade noticeably when you drop from frontier models to a 32B-parameter local checkpoint. But for read-heavy work (code review, explanation, single-function rewrites), and especially for codebases under NDA or air-gap, a local sidebar wired to a decent local model is now a credible default.

The broader stack implication: expect a Cambrian explosion of single-purpose MCP servers over the next quarter, each picking off one piece of the cloud-managed agent loop. Project memory (Recall) is the first. Tool routing, prompt history, code-search indices, and per-repo policy configs are obvious next targets. The interesting question isn't whether these get built — they will — but whether Anthropic and OpenAI respond by opening their managed equivalents or by tightening the moat.

Looking ahead

The arc here is familiar. Every cloud category eventually grows a local-first counter-stack, and the counter-stack wins on three predictable axes: data ownership, cost at scale, and the long tail of regulated or air-gapped users who can't use the cloud version at all. The AI coding tool market is roughly 18 months into its hype cycle, and the local-first defection is starting on schedule. Recall and Junction are early enough to be rough; they're also early enough that installing them this week buys you a year of compounding context that the cloud-only crowd will be re-creating from scratch when their vendor pivots or raises prices.