syrup
A coding agent in your browser, powered by whatever API keys you already have. Run it on your own machine (this README), or use the hosted version at https://syrup.syedsarib.com (sign in with Google; see docs/PLAN.md).
No subscription. No single vendor. Start with free models that need no key, then bring keys from NVIDIA, OpenRouter, Mistral, Z.ai, Google AI Studio, or any of 200+ providers. syrup routes work across your keys, fails over when a free tier hits its limit, tracks every token, and shows you exactly what each session cost.
What you get
- A full coding agent in a clean chat UI: streaming replies, tool calls (read, write, edit, run, search), permission prompts, plan/build agents, subagents, LSP diagnostics, context compaction. The engine is OpenCode, the most-used open-source coding agent.
- Workspaces. Everyone starts in Home, a folder syrup makes for you (
~/syrup, %USERPROFILE%\syrup on Windows), so there is nothing to pick before the first chat. Add any other folder with the native OS folder dialog, or type a path. Chats are per folder.
- Bring your own keys. Add as many keys as you like per provider, label them, mark them free or paid. Keys are encrypted at rest and never leave your machine.
- Router. Auto picks the best model for each chat from your keys. It sticks with that model for the rest of the chat, so prompt caches stay warm and answers stay consistent. If a model is rate-limited, overloaded or slow to start, syrup switches to the next one before any text reaches you, so you do not see the error. Fast does the same with quicker models and low thinking effort, for small tasks.
- Model picker. Auto and Fast come first, then your favorites, recent models and recommended ones. Everything else sits behind All models.
- Cost dashboard. Tokens and dollars per message, session, model and day. Free-tier usage is shown separately from real spend.
- Long-term memory. The agent saves and searches facts across sessions (SQLite + full-text search), exposed to it as MCP tools. You can edit everything it remembers.
- Skills. Install
SKILL.md skills from GitHub or paste your own. Works with skills you already have in ~/.claude/skills.
- Attachments. Drop or paste images and files into the composer.
- Logs. A structured application log (secrets redacted), one click away in the sidebar, with a copy-for-debugging button.
- Optional
.sarib tools. In workspaces that contain .sarib files, the agent can query and edit them by id instead of rewriting them. Other workspaces pay nothing for it. One click on the Skills page installs it.
- Free out of the box. OpenCode Zen's free models work with no key at all, so you can try it before adding anything. Pick one directly in the model picker; Auto and Fast need at least one key.
Requirements
- Node 22+ (24 recommended) and pnpm 10+
- The local server listens on
127.0.0.1 only. Provider keys are encrypted with a key kept in %APPDATA%\syrup (Windows) or ~/.config/syrup. See SECURITY.md.
- Git (used to install skills from GitHub)
- OpenCode installed globally:
npm i -g opencode-ai
- Linux only:
zenity or kdialog for the folder dialog (you can always type a path instead)
- Optional: Python 3.10+ for the
.sarib tools (the Skills page installs the rest)
Run it
git clone https://github.com/SyedSaribSultan/syrup.git
cd syrup
pnpm install
cp .env.example .env.local # optional; defaults work
pnpm dev # http://localhost:3000
For a production build:
pnpm build
pnpm start
On first launch syrup starts an OpenCode server, the router and the memory server for you. Nothing else to configure.
First steps
- Open Providers and add a key. See Which keys to add first. Auto and Fast route only across your keys.
- Start in Home, or pick another folder with the workspace switcher at the top of the sidebar.
- Start a chat with Auto selected. No key yet? Pick an OpenCode Zen free model in the model picker instead.
- Check Usage & cost afterwards to see what it used.
How it fits together
Browser (Next.js UI)
│ /api/oc/* (typed OpenCode SDK through a proxy)
▼
syrup server ── router :4210/v1 ── provider APIs (your keys)
│ memory :4210/mcp
│ SQLite data/syrup.db
▼
OpenCode server :4096 (agent loop, tools, LSP, MCP, skills)
See docs/ARCHITECTURE.md for the full picture and the reasoning behind it.
Configuration
Everything is optional. Copy .env.example to .env.local to change:
| Variable | Default | What it does |
|---|
OPENCODE_PORT | 4096 | Port for the embedded OpenCode server |
OPENCODE_URL | – | Attach to an OpenCode server you run yourself instead of spawning one |
SYRUP_ROUTER_PORT | 4210 | Port for the router and memory MCP server |
SYRUP_WORKSPACE | cwd | Folder the engine starts in. The app opens Home (~/syrup) on a first visit and uses this only if Home cannot be created |
SYRUP_DB | file:./data/syrup.db | SQLite location |
SYRUP_VAULT_KEY | auto-generated | 32-byte base64 key that encrypts provider keys. Generated into data/vault.key if unset |
Provider keys can also come from the environment (GOOGLE_API_KEY, GROQ_API_KEY, MISTRAL_API_KEY, OPENROUTER_API_KEY, …). The router uses them when no key is saved in the vault.
Free tiers, honestly
Free tiers change often. As of September 2026:
- OpenCode Zen — free models with no key at all (Big Pickle, Muse Spark 1.3, MiMo-V2.6-Flash, Space Bunny and more). Pick them directly in the model picker. Auto and Fast can't route them, because OpenCode limits its free tier to OpenCode itself. Free for a limited time; some use your prompts for training.
- Google AI Studio — best free quality: Gemini 3.8 Flash with 1M context. But the free tier is small: about 20 requests/day for Flash, about 500/day for Flash-Lite. Quotas are per model and reset at midnight Pacific. Pro models are paid only. Prompts may be used for training.
- NVIDIA NIM — the largest free volume: about 40 requests/minute, no published daily cap. Free Kimi K3, GLM-5.3, DeepSeek V4 and Nemotron. Terms say development and evaluation use.
- OpenRouter — one key, 20+ free models. 50 free requests/day shared across all free models; 1,000/day after a one-time $10 purchase.
- Mistral — free plan with $10 of API credits every month, no card (phone verification). Devstral, Medium 3.5, Codestral. Turn off training in Privacy settings.
- Z.ai — GLM-4.7-Flash is free with no daily cap, one request at a time.
- Cohere — trial key with 1,000 calls/month. Non-commercial use.
- Groq — very fast, but the free tier allows 8K tokens/minute. One agent turn is often bigger than that, so syrup uses Groq only for quick side tasks.
- Cerebras — no free tier any more: a $5 trial with a card, 30K tokens/minute.
Which keys to add first
OpenCode Zen works with no key, so you can start right away by picking one of its models directly. Auto and Fast need keys. Add them in this order:
- NVIDIA NIM — the most free requests. https://build.nvidia.com/settings/api-keys
- OpenRouter — many free models behind one key. https://openrouter.ai/settings/keys
- Mistral — monthly free credits and good coding models. https://console.mistral.ai/api-keys
- Z.ai — a free GLM model with no daily cap. https://z.ai/manage-apikey/apikey-list
- Cohere — a small monthly trial allowance. https://dashboard.cohere.com/api-keys
A Google AI Studio key is still worth adding for its quality, but it runs out quickly on the free tier. The Providers page shows the same guidance next to each provider, and you can tag each key free or paid so the router knows what it costs.
Sharing chats
Press Share in a chat's header to get a link to a read-only snapshot of it. Anyone with the link can open it without an account; later messages stay private until you Update link, and Stop sharing takes it down within seconds. Snapshots include everything (messages, tool steps with inputs and outputs, reasoning, small attachments); things that look like secrets are redacted and file paths are shortened. Shared pages are never indexed by search engines.
Every link also has machine-readable versions, handy for debugging with other tools and agents:
/c/<id>/md (or /c/<id>.md): a clean Markdown transcript/c/<id>/json (or /c/<id>.json): the structured syrup.transcript v1 format (see src/lib/transcript.ts)/c/<id> with Accept: text/markdown or application/json returns the same- Copy debug link (owner only, expires after 24 hours) adds the router's record of each step: which model answered, retries, errors and timings
Locally, a link only opens on your machine, so the Share panel also exports Markdown, JSON or a self-contained HTML file. From a terminal, pnpm chat:export <sessionId|latest> [--json] [--debug] prints a local chat the same way.
Hosted version
The same codebase runs as a hosted product at https://syrup.syedsarib.com, open to anyone with a Google account: Google sign-in, per-user encrypted keys, and each workspace's agent in its own isolated sandbox. Design and roadmap: docs/PLAN.md, docs/PHASE2.md. To run your own hosted instance, follow docs/DEPLOY.md.
Status
Working end to end, early. Things still rough: no theme toggle, no mobile layout, the router has been tested against real provider endpoints but not yet under sustained rate limiting. Issues and PRs welcome.
Author
Built by Syed Sarib Sultan, senior product design engineer. More work and guides at syedsarib.com.
License
MIT
