Introduction
stella is a fast, bring-your-own-key, model-agnostic terminal coding agent. Point it at any provider, give it a goal, and let a verifier decide when done.
stella is a coding agent that runs in your terminal and works with any model provider. It's bring-your-own-key: there's no account or sign-up for Community use, it runs on credentials you already have, it records everything locally, and it sends telemetry nowhere unless you turn that on yourself.
export ANTHROPIC_API_KEY=your_key_here
stella run "add a --json flag to the export command and update the tests"What it does differently
| BYOK, no lock-in | Finds your provider automatically from whatever keys you have set: Anthropic, OpenAI, Gemini, Vertex, Bedrock, xAI, DeepSeek, Z.ai, OpenRouter, or any OpenAI-compatible local server. |
| Verified outcomes | Goal mode doesn't stop on a guess. A separate verifier model checks the evidence against the definition of done before the loop ends. |
| Tools behind a permission model | A small built-in toolset — the shell, file editing, code search, sub-agents, the task board, scratch state, and an environment check — plus MCP servers and your own script tools, each one gated with optional lifecycle hooks. |
| Config that scales to a team | Set a different base URL, key, or model per provider in settings.json, merged across a user, org, and project hierarchy. |
| Local-first telemetry | Usage, cost, and per-step metering land in .stella/private/store.db on your disk. |
Bring your own provider
stella speaks each provider's own protocol instead of forcing every call through one OpenAI-shaped adapter. That means provider features like Anthropic's thinking blocks, Gemini's thinking levels, and Bedrock's Converse format work natively, not as an imitation.
How it fits together
stella sends your prompt to the model you picked. The model calls tools to
read and change your workspace, stella feeds the results back, and this
repeats until the task is done. Behind the scenes: the tool registry, MCP
servers from .stella/mcp.toml, and the memory and code-graph context
engine. In goal mode, a second
model checks the evidence, so the loop ends on a verified result instead of
the worker's own claim that it's done.
Common tasks
Each of these walks through one real task to a finished, verified result. They link down to the full reference instead of repeating it, so you only read the complete flag list when you actually need it.
Watch a red branch, fix the root cause, push, and check again. Stop only when the run is green for everyone, not just you.
Use the code graph to see what a change would affect before you commit to an approach. Then ship it verified.
From the total dollar cost down to the exact bytes sent in one model call, all stored on your own machine.
Set a hard spending ceiling, see where a run stops if it hits that ceiling, and understand what happens when the transcript outgrows the model's context window.
New to stella? Get started takes about thirty
seconds: install, set one key, run stella init, and make your first
verified change. Something not working?
Troubleshooting is organized by symptom.
Where to go
| Section | Read it for |
|---|---|
| Start | Installing stella, setting up a provider key, starting a project, and making your first verified change. |
| Do | Step-by-step guides for real tasks: fixing red CI, working in an unfamiliar repo, tracking costs, setting budgets, fleets, headless scripting, and embedding the engine in your app. |
| Understand | How plugins prove work is done, how the engine's loop works, how the context engine keeps recall cheap, the modes you can run stella in, and where telemetry goes. |
| Configure | settings.json, providers, tools and permissions, and ready-made recipes. |
| Reference | Every command and flag, the event bus, and release notes. |
| Project | The showcase, and supporting stella. |
stella never routes model traffic through a hosted service of its own. Your keys call your provider directly. Community telemetry stays on your machine. Only an Oxagen Enterprise seat you've explicitly enrolled can send a small, content-free operational summary to one approved address. See Telemetry.