I only found this discussion: https://github.com/zed-industries/zed/discussions/24004 and find it a bit surprising that this is not yet an implemented feature. Literally every other IDE/Editor I know has this.

Hi. I'm not really knowledgeable with rust in any facet. So, funny enough; I had Zed vibe code Antigravity provider support for itself. I used this opencode plugin as reference, and at least from my tests (~2-3 hours of total work); all the models seem to work fine.

It can be added from the same LLM providers page in settings, I neglected to show it because I do not want to show my linked email(s).

If this is something that interests you, here is my fork of the project. Star it maybe, numbers make me happy.
https://github.com/kevinbudz/zed-agy

I'll try and update it semi-regularly in the event of file conflicts. I will also update this post, and probably make a new post if I add support for Cursor, too. I am aware it has ACP support and you can open up a 'New Cursor Thread'. But, that option also prevents you from using models from other providers if you were to, let's say- use 3.5 Flash on your Antigravity sub for planning, then something like Composer 2.5 for acting, etc..

Anyways, bye now!

—- !!!
Integrations like this are against Google’s TOS and can get you blacklisted from Antigravity.

(I have been using the OpenCode plugin this project is based off of for several months without issue. But, that doesn’t mean it’s 100% safe)

So I'm trying to set up a very simple agent with the model running using llama-server but I keep running into this "Failed to write API key to keychain" error every single time.

I tried adding the keys in system environment variables (OPENAI_API_KEY, AMAZING_LOCAL_PROVIDER_API_KEY) but none of them seem to bypass that error. I configured the prediction part which is picking up the API Key properly but it doesn't predict anything, the requests don't actually seem to reach the server.

I verified the server is working fine by heading to localhost:8080 and chatting there.

Here's the command for llama-server if that's relevant

.\llama-server.exe -m "C:\Misc\Models\Ornith-1.0-9B-MTP-Q4_K_M.gguf" --n-gpu-layers 99 --ctx-size 8192 --flash-attn on --jinja --spec-type draft-mtp --spec-draft-n-max 3

I tried with Ollama and that's working fine, but I am more interested in running it via OpenAI Compatible API.

The checkboxes checked (cropped in the image) are Supports tools and Supports thinking. Tried unchecking and stuff but no dice.

RESOVLED: Credential Manager hit a limit in the no. of credentials it can store. Deleted one old useless cred and tried it again, works fine now.

Most of my workflow already lives in Neovim — code, prose, notes, scratchpads. The piece that always lagged was querying the notes. Plenty of tools let me grep them; almost none let me ask things like "all the drafts under tasks/q2 that link to people/alice" without leaving the buffer.

Turns out you can. IWE is a Rust binary (LSP server + CLI) that treats a directory of .md files as a queryable graph. Install once, use it from the editor over LSP and from the shell over :!.

The query language is small and reads like Mongo's:

iwe find --filter 'status: draft, priority: {$gte: 8}'

iwe find --filter 'author.email: {$exists: true}'

Frontmatter is the schema. Markdown links are the relationships — and there are two kinds, which the engine actually distinguishes:

• An inline link in body text is a reference: "see also."
• A markdown link alone on its own line is an inclusion link: containment. The linked document becomes a structural child of this one.

Each gets its own pair of operators:

iwe find --references people/alice # docs that link to Alice inline
iwe find --included-by tasks/alpha:0 # everything under alpha's tree (unbounded)
iwe find --included-by tasks/alpha:0 --references people/dmytro --filter 'status: draft'

That last line: drafts under the tasks/alpha subtree that also mention people/dmytro inline. Three relationships, three flags.

The same predicates drive iwe count, iwe update, iwe delete. Bulk-set frontmatter from the shell:

iwe update --filter 'status: draft, reviewed: true' \
--set status=published \
--set published_at=2026-05-02

update and delete require an explicit --filter (no accidental whole-corpus rewrites). --dry-run previews.

From inside Neovim, this composes two ways.

The same iwe binary is also a markdown LSP server, so the editing side feels like working in code:

• gd — jump to linked notes
• gr — find references / backlinks
• K — hover preview of a linked note without opening it
• Code actions for extract section to a new file, inline a referenced note, rename
• Auto-complete for link targets as you type
• Inlay hints showing parent context and link counts

There's a dedicated plugin — iwe.nvim — that wires the LSP up and adds Telescope integration with hierarchical path search (notes show as Journal ⇒ 2026 ⇒ Week 18 ⇒ Mon notes). Lazy / packer / vim-plug all work.

For querying, you don't need a special integration — the CLI is enough:

• Output is plain text — pipe to jq, fzf, telescope, whatever.

Same install handles both: cargo install iwe and you have the LSP server + the CLI. The LSP runs against any folder of .md files; the CLI queries the same folder.

Side note: this also turns out to be a decent shape for AI agents. They use the same CLI you do, see the same files, and git log is your audit trail for whatever they touch.

Repo: https://github.com/iwe-org/iwe · Plugin: https://github.com/iwe-org/iwe.nvim

Curious what the heavy notes-in-Neovim crowd thinks, especially on the inclusion-vs-reference link split.

I have been a subscriber to Jetbrains since last 8 years. And I always wanted to have an alternate solution. Neovim/Lazyvim made me confident but that's a TUI that needs quite a bit of config before all pieces work together. I needed something that was straight up easy, just as performant (my main pet peeve with Jetbrains tools other than price), supported debugging, multiple panes, easy enough UI etc.

My projects are usually monoliths and memory consumption in VS code made my laptop cry.

Zed fixed all of those. Every single thing except the easy "refactor" in Goland, which I had stopped using because I got better at designing things from the ground up.

So again, thanks for the editor and thank you for keeping it free.

Edit: My card got renewed, and I knew it would fail automatically, so didn't bother cancelling the subscription manually.