opencode kit

Portable global opencode setup: one non-vision chat model (Nemotron) that can still see (vision via NVIDIA API), imagine (FLUX via Cloudflare), and drive devices/browsers/design/docs through MCP servers — in every project, on every machine.

What's inside

| Path | Purpose | |------|---------| | opencode.jsonc | Global config: model pin, provider whitelist, MCP servers (execbro, playwright, stitch, context7, discord-agent). Secrets referenced as {env:...} — never stored here. | | tool/analyze_image.ts | Vision for the chat model — sends any image to an NVIDIA vision model, returns text. Auto-compresses to the API limit. | | tool/generate_image.ts | Image generation — Cloudflare Workers AI FLUX.1 Schnell, Pollinations fallback. | | lib/vision.ts | Shared NVIDIA vision client used by the tools. | | skill/vision-handoff.md | Teaches the model when/how to use the tools (and to use execbro/playwright for screenshots instead of duplicating them). | | agent/coder.md | Generic coder agent with the full tool roster. | | setup.sh | Interactive setup: dependency install + paste keys + shell wiring + health checks. | | secrets.env | Your keys (gitignored, chmod 600). Created by setup.sh. |

New machine (cross-system)

One-liner (npx):

npx @scripting-bear/opencode-kit

That downloads the published kit, copies it into ~/.config/opencode (asks before overwriting; never touches secrets.env), installs tool dependencies, and drops you into the interactive key setup. Flags: --yes (no overwrite prompt), --no-setup (skip keys). New versions: bump version in package.json and npm publish from the source checkout.

Or via git (keeps the machine syncable with git pull):

git clone <your-remote> ~/.config/opencode
~/.config/opencode/setup.sh     # interactive: checks deps, prompts for keys
# open a new terminal, then:
opencode mcp list               # every server should show ✓ connected

That's the whole migration. Projects need no per-project setup; a project's own opencode.json / .opencode/ merges on top of this if present.

Managing keys & customizations

• TUI/CLI: ~/.config/opencode/setup.sh — re-run any time; Enter keeps existing values, paste to replace.
• Editor ("GUI"): edit ~/.config/opencode/secrets.env directly, then open a new terminal.
• opencode's own TUI: opencode auth login for chat-provider credentials (stored in ~/.local/share/opencode/auth.json, separate from this repo); /models in the TUI to switch models.
• Customizations: edit opencode.jsonc (MCP servers, model), skill/, agent/, tool/ — then commit & push; other machines git pull.

Keys used

| Env var | Used by | Where to get it | |---------|---------|-----------------| | NVIDIA_API_KEY | analyze_image (vision) | build.nvidia.com | | STITCH_API_KEY | stitch MCP | stitch.withgoogle.com | | CONTEXT7_API_KEY | context7 MCP | context7.com/dashboard | | CLOUDFLARE_API_TOKEN | generate_image | Cloudflare dashboard → API tokens | | CLOUDFLARE_ACCOUNT_ID | generate_image | Cloudflare dashboard / npx wrangler whoami |

Notes

• secrets.env is sourced from your shell rc (setup.sh adds the line), so keys reach both opencode's {env:...} config substitution and the custom tools' process.env.
• mcp.discord-agent points at a local build (machine-specific path) — fix the path or set "enabled": false on machines without it.
• Vision image compression uses macOS sips; on Linux, swap it for ImageMagick in lib/vision.ts if you ever need it there.