Noob
Small fixes.
Light experiments.
Enough to feel Dragon wake up.
[ Dragon Code ]
Frontier coding, off the leash.
A terminal agent built on trillion-parameter models with a real 1B-token addressable context. It reads your whole repo, edits real files, and cites every change to file:line.
No five-hour windows. No rolling lockouts.
Work until you're done.
[ Get started ]
Install Dragon Code
Copy this, paste it into your terminal, and press Enter. On Windows, use the PowerShell command from the install guide.
curl -fsSL https://raw.githubusercontent.com/VELLORAAI/dragoncode-public-dist/main/install | bash[ Example session ]
Grounded, not guessing.
dragon · example session
$ dragon
Indexed 4,812 files · 1B tokens addressable
> the checkout webhook double-charges customers sometimes. why?
Reading api/webhooks/checkout.ts:118-164
Reading lib/queue/consumer.ts:42-90
checkout.ts:142 processes the event before consumer.ts:77 writes the idempotency key. Two webhooks landing ~40ms apart both pass the check.
Moving the idempotency write before processing.
✓ api/webhooks/checkout.ts:142 edited
✓ tests/webhooks/checkout.test.ts added
✓ 12/12 tests passing
[ Dragon Code ]
Drop into any repo. It already has the whole thing in context.
Install the dragoncommand, paste your key, go. It indexes your repo once, then keeps the right code loaded as you work, so every answer is grounded in what's actually there, not a guess.
How Dragon reads your whole repo →
Every claim cited to file:line, or it says it isn't sure.
Real code search, not blind grep.
Remembers your project across sessions. Reads your AGENTS.md / CLAUDE.md.
Same setup in any repo. Paste a key and go.
macOS, Linux, and Windows.
[ Dragon ]
Trillion-parameter models, running unrestricted.
Trillion-parameter open-weight models, tuned for code. Frontier capability without the frontier lab's price or its limits. Default to the Dragon lane, drop to /deepseek-flash for cheap work, step up to /k2.7, or hit /glm-5.2 or /k3 for the hard problems. Every lane is named after the model that serves it.
Trillion-parameter open models, tuned for code.
Frontier capability on the default lane. /deepseek-flash, /k2.7 or /glm-5.2 to switch.
No mid-session downgrades to a cheaper model.
A fraction of proprietary cost, at full capability.
Prefer raw inference? Same models, drop-in API →
[ Advisor ]
A frontier model on call.
Advisor runs automatically. The fast model keeps driving your session, and when it hits a call worth getting right, it hands that one question to a flagship frontier model, takes the answer, and keeps building. You pay frontier rates for that single call, not your whole session.
Everyday work
The fast model drives your session.
Hard call
It asks a flagship frontier model, one question.
Back to work
Takes the answer, keeps building.
[ Privacy ]
Your index. Your key. Your call.
Dragon Code indexes your repo so retrieval works. That's the only reason it uploads anything.
Your index is owner-scoped. Only your key can read it.
Secret files like .env, keys, and credentials are filtered out automatically, before upload and again on our end.
Turn indexing off anytime, or delete your index whenever you want.
Not "trust us."
We never train on your code.
[ Pricing ]
Buy runway. Not permission.
Monthly Dragon token pools that scale with how much you actually build.
No surprise bills. Run past your pool, top up with credits anytime.
No lock-in. Cancel or change plans before your next renewal.
Starter
Daily edits.
Small features.
Bug fixes.
Actual usage without instantly watching the meter.
Builder
Big refactors.
Agent runs.
Longer sessions.
Enough runway to stay in motion.
Popular
Pro
Deep sessions.
Large repos.
High-stakes builds.
The plan for people who use AI coding like oxygen.
Supermax
Maximum runway.
Maximum throughput.
Deep, all-day agent runs across large codebases.
For teams and heavy builders who live in the terminal.
What tokens cost per million
Tokens are metered against your monthly allowance at these per-million rates: an input rate, an output rate, and a cached-input rate per model. Reasoning tokens bill as output.
| Model | Best for | In / 1M | Out / 1M | Cached in / 1M | SWE-bench Verified | SWE-bench Pro |
|---|---|---|---|---|---|---|
| Dragon | Everyday coding | $0.1294 | $0.9412 | — | 76.2% | 50.9% |
| DeepSeek V4 Flash | Cheapest lane, 1M context (/deepseek-flash) | $0.1647 | $0.3294 | $0.0329 | — | — |
| Kimi K2.7-Code | Harder everyday tasks (/k2.7) | $1.1176 | $4.7059 | $0.2235 | 78.2% | 58.6% |
| GLM-5.2 | Hardest tasks (/glm-5.2, or append /max) | $1.6471 | $5.1765 | $0.1647 | — | 62.1% |
| Kimi K3 | Frontier lane, 1M context (/k3) | $3.5294 | $17.6471 | $0.3529 | — | — |
| DeepSeek V4 Pro | Frontier agentic lane, 1M context (/deepseek-pro) | $1.5529 | $4.6588 | $0.0518 | — | — |
| Qwen3.8 Max | Long-horizon agentic tasks (/qwen3.8) | $2.3529 | $7.0588 | $0.2941 | — | — |
| MiniMax M3 | Cheapest long-context lane, 512k (/minimax-m3) | $0.3529 | $1.4118 | $0.0694 | — | — |
Cached in / 1M is the rate for input tokens served from a warm prompt cache; it applies automatically when a turn reuses an identical prefix. “—” in that column means the lane passes no cached discount. SWE-bench scores are each model provider’s own published Pass@1, not independently re-run by us. “—” there means the provider hasn’t reported that suite.
FAQ
Common questions.
Dragon Code is a terminal-native AI coding agent (the `dragon` command) wired to Dragon inference at api.mws.run. It runs trillion-parameter models with a 1B-token addressable context and indexes your whole repo, pulling the most relevant code in on every turn. Every claim is grounded in real file:line references.
[ Get started ]
Put a frontier agent on your codebase.
The same gateway has run Dragon Code and our own workloads in production for months. Grab a key, install the dragon command, start building.