AIDE
The AI Development Engine: a terminal coding agent on a provider-agnostic LLM gateway.
One engine that plans on the strongest model, implements on the most economical one, falls back automatically when providers throttle, and runs just as well against your own GPUs as against any cloud.
Every phase on the right model
AIDE splits software work into phases and routes each one to the model tier where it is most cost-effective: planning and review on frontier models, implementation on balanced ones, tests and docs on efficient ones. A budget estimator prices the whole job first, shows the savings against running everything on the frontier tier, and asks before spending.
One gateway, every provider
Anthropic, OpenAI, Azure, Google, DeepSeek, and self-hosted backends sit behind a single gateway: one place for authentication, rate limits, cost accounting, and model governance. When a provider throttles, work cascades down a configured fallback chain instead of stopping.
Local models, local code
Point the gateway at models running on your own GPUs and 70B-class open models serve any phase. Code and prompts never have to leave your machines, which is the difference between adopting AI tooling and getting it past security review.
Retrieval over a code knowledge graph
AIDE builds a local knowledge graph of the repository from syntax trees, spending zero model tokens on parsing. Questions about the codebase resolve through graph queries instead of file reading, refactors get a verified blast-radius analysis before any edit, and useful answers persist so knowledge accumulates across sessions.
Terminal, server, desktop
The same engine runs three ways: a full terminal UI for daily work, an HTTP server for remote setups, and a cross-platform desktop client. Run the models on a GPU server, drive them from a laptop.
Thirty bundled skills, MCP native
Code review, security audit, test generation, refactoring, documentation, dashboards, diagrams, and embedded-systems checks (MISRA C and C++, Verilog linting) ship as skills that trigger by command or keyword. MCP support connects any external toolserver; subagents fan work out under the same budget plan.
Where each phase runs, and why
| PHASE | TIER | WHY |
|---|---|---|
| Plan | Frontier | Architecture decisions need the deepest reasoning |
| Review | Frontier | Adversarial review catches what implementation misses |
| Analyze | Balanced | High-volume code reading at the best quality per dollar |
| Implement | Balanced | Bulk code writing with strong quality |
| Test | Efficient | Mostly mechanical; a small model matches frontier quality |
| Document | Efficient | Same economics as testing |
Every mapping is configuration, not code: presets range from everything-on-one-provider to fully open-source, and any gateway backend can serve any phase. Pin a phase to an exact model, replace a tier's fallback chain, or disable routing entirely with one setting.
Private preview
AIDE is in private preview inside Proxiant Labs, running daily production work across our own products. If phase-routed orchestration on your own hardware is the shape of tooling your team needs, talk to us.