You've been feeding Cursor $20 to $60 a month since last winter, squinting at the "memory lock-in" complaints piling up on forums through early April 2026, and quietly wondering whether the only exit is crawling back to VS Code with a pile of extensions duct-taped together. The subscription is not the problem. The switching cost is. Every prompt you wrote, every rule you tuned, every correction the agent learned from — that context lives inside a closed box, and when Cursor killed its Memory feature in v2.1.x and told users to "export to Rules," the export path was, let's say, aspirational.

This is the familiar AI-editor trap. Every credible tool in this category has been closed-source, token-hungry (an LLM — large language model, the brain behind ChatGPT and Claude — charges by how much text it reads), and quietly hoarding context you can never walk out with. The lock-in is no longer the price. It's the memory.

Then Zed — a Rust-based code editor, open source under GPL v3 — spent the last two weeks quietly calling the bluff. Between April 1 and April 15, 2026, the team shipped five releases (v0.230.0 through v0.232.2) and a preview build (v0.233.0-pre). The stack now includes a full Agent Panel with a native MCP (Model Context Protocol — a universal plug standard that lets AI tools talk to each other, like USB for data) client, multi-model routing across Claude, GPT, Gemini, Bedrock, and local models via Ollama, plus inline edit prediction. Bring your own API key (API — how programs talk to each other) and the incremental cost is zero 😼.

The April cadence is the actual story. v0.231.1 on April 8 added top-down streaming for agent threads and a native devcontainer implementation. On April 10, Zed launched Agent Metrics — a public weekly dashboard showing session counts, turn volume, and model latency. It immediately exposed that Claude Sonnet's p90 latency climbed 44% over three weeks, a number no closed editor would ever publish. April 15's v0.232.2 added nine Bedrock models (NVIDIA, Z.AI, Mistral, MiniMax) and rewrote agent panel error messages into actual English. Zed's own numbers: 0.4s startup, 2ms input latency — versus Cursor's ~12ms on its Electron base (Electron — the tech that lets web code pretend to be a desktop app, usually poorly).

Here's the uncomfortable read: the AI editor is now a commodity wrapper 😹. The agent loop, the tool protocol, the multi-model router — all open-sourced. What Cursor actually sold for $2B ARR was UX polish and distribution. The irreplaceable asset was never the autocomplete. It was the accumulated memory they refuse to let you export.

The tradeoffs are ugly and honest. Zed has no enterprise SSO story (SSO — one login for all your work tools), no sandbox-isolated cloud VMs like Cursor 3 ships, a thinner extension ecosystem, and no GA Windows build. Model quality still depends on whichever API key you wire in, and "free" means you do the operational work Cursor hides behind the subscription 😾.

For you, the question just flipped. It's no longer "which AI editor is smartest." It's "which one lets me leave with my context intact." For the first time since 2023, procurement has a credible open-source answer to point at.

The coding-tools market got its escape hatch. Every closed AI editor now has to justify the lock-in on something other than "we have the agent panel" — because that's Apache 2.0 now 🐈‍⬛.