Runs inside OpenCode
The installer detects your OpenCode major version and copies the matching plugin. On OpenCode 2 the zed-ai provider registers through the in process plugin API, so no extra process is needed.
v2.0.0
Use your Zed plan
from any CLI agent
without the editor
zed2cli bridges the AI models in your Zed account to OpenCode, pi, Hermes, and any tool that speaks the OpenAI API. It runs on your machine and talks to cloud.zed.dev for you.
$npm install -g zed2cli
LIVE BRIDGEcloud.zed.dev/completions
In-Process Hook
AGENT:
01 CLIENTMemory IPC (no proxy)
> opencode run -m zed-ai/claude-sonnet-4-6 "optimize parser"
OpenAI Spec ⇄ Zed Protocol
zed2cli engine Keyring Auth ~50m
JWT Auto-Mint SSE ⟷ NDJSON 0 Runtime Deps
Cloudflare Edge (cf-ray verified)
03 ZED CLOUDHTTP 200 STREAM
01event: message_start (role: assistant)
█
16Zed Models
0Runtime Deps
50Min Token Refresh
How the bridge works
Zed speaks its own protocol. zed2cli translates both directions. These are the files a contributor would touch.
OpenCode talks to the layer in process. pi, Hermes, and other clients talk to it over HTTP on 127.0.0.1:8787. Both leave through the same translation code.
PLUGIN
OpenCode plugin
src/index.mjs registers the zed-ai provider through the OpenCode 2 plugin API and rewrites requests in process. src/index.v1.mjs is the OpenCode 1 build, and the installer copies whichever one matches your version.
BRIDGE
Proxy and pi
src/proxy.mjs serves GET /v1/models and POST /v1/chat/completions, translates each request, and forwards it to cloud.zed.dev. src/pi/zed-ai.mjs points pi at that proxy through ZED_PROXY_URL.
CATALOG
Shared model catalog
src/models.mjs is the single source of truth for the 16 models, their provider keys, and their capability flags. The plugin and the proxy both import it.
TOKENS
Token resolution
scripts/zed-token.sh resolves a JWT from ZED_TOKEN, the system keyring, ~/.zed_access_token, or ~/.zed_jwt_token. The refresh daemon keeps the file current so long runs do not expire.
From install to a running agent
Real commands and real output. Each tab is a sample run on a machine with Zed signed in and the tool already installed.
opencode, plugin install
$ npm install -g zed2cli
$ zed2cli install-opencode
detected OpenCode 2.0.29
installed the plugin for the v2 API
$ opencode models | grep zed-ai
zed-ai/claude-sonnet-4-6
zed-ai/gpt-5.6-sol
zed-ai/gemini-3.5-flash
... 13 more
$ opencode run -m zed-ai/claude-sonnet-4-6 "say hello in one word"
helloCapabilities
Everything zed2cli wires up between install and a working agent. Search by keyword, filter by scope.
Local OpenAI bridge
Point pi, Hermes, or any OpenAI compatible client at http://127.0.0.1:8787/v1. The proxy rewrites each request into the Zed envelope and streams the answer back as server sent events.
pi provider in one command
zed2cli install-pi registers the zed-ai provider with pi and points it at the local proxy. The extension carries its own model list because pi loads extensions from a virtual URL.
Tokens stay fresh
A Zed JWT lasts about an hour. The refresh daemon keeps a live token in ~/.zed_jwt_token every 50 minutes, and every command, the plugin, and the proxy read it from there.
doctor tells the truth
One command lists every token source with its expiry, checks whether the plugin matches your OpenCode version, and reports whether the proxy is running.
One source of truth
src/models.mjs lists all 16 models with their provider key, context size, output size, and capability flags. The plugin and the proxy both import it, and a check fails the build when the pi copy drifts.
Models in the Zed Catalog
Every model below was read from GET https://cloud.zed.dev/models with a Zed Student plan token. Other plans may expose a different list.
| Model | Provider | Context | Max Output | Reasoning |
|---|---|---|---|---|
| Claude Sonnet 5 | Anthropic | 1,000,000 | 128,000 | yes |
| Claude Sonnet 4.6DEFAULT | Anthropic | 1,000,000 | 64,000 | yes |
| Claude Sonnet 4.5 | Anthropic | 200,000 | 8,192 | yes |
| Claude Haiku 4.5 | Anthropic | 200,000 | 64,000 | yes |
| GPT-5.6 Sol | OpenAI | 400,000 | 128,000 | yes |
| GPT-5.6 Terra | OpenAI | 400,000 | 128,000 | yes |
| GPT-5.6 Luna | OpenAI | 400,000 | 128,000 | yes |
| GPT-5.5 | OpenAI | 400,000 | 128,000 | yes |
| GPT-5.4 | OpenAI | 400,000 | 128,000 | yes |
| GPT-5.3-Codex | OpenAI | 400,000 | 128,000 | yes |
| GPT-5.2 | OpenAI | 400,000 | 128,000 | no |
| GPT-5 mini | OpenAI | 400,000 | 128,000 | no |
| GPT-5 nano | OpenAI | 400,000 | 128,000 | no |
| Gemini 3.1 Pro | 200,000 | 65,535 | yes | |
| Gemini 3.5 Flash | 1,048,576 | 65,535 | yes | |
| Gemini 3 Flash | 1,048,576 | 65,535 | yes |
All 16 models support tools and images. Reasoning is the supportsReasoning flag from src/models.mjs. The default model is claude-sonnet-4-6.Catalog Verified
Keep the bridge open
zed2cli is MIT licensed and built in the open. If it saved you from writing a proxy of your own, star the repo or pass it on to another engineer.