Introduction

Stella is a fast, model-agnostic coding agent that lives in your terminal. It is bring-your-own-key (BYOK): there is no account and no sign-up — Stella runs on the provider credentials you already have, and everything it records stays on your machine.

export ANTHROPIC_API_KEY=your_key_here
stella run "add a --json flag to the export command and update the tests"

What makes Stella different

• BYOK, no lock-in. Stella auto-detects the provider from whichever API keys you have set. Bring 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 hunch — a separate judge model verifies the definition of done from evidence before the loop ends.
• Real tools with a permission model. Read, edit, glob, grep, shell, CI, screenshots, issue tracking, and developer-defined script tools — each gated by a per-tool permission model and optional lifecycle hooks.
• Configuration that scales to a team. Override any provider's base URL, key, or model in settings.json, merged across a project → org-managed → user hierarchy.
• Local-first telemetry. Token usage, cost, and per-step metering land in a local SQLite store (.stella/store.db) on your disk. Nothing is sent anywhere.

How it fits together

  you ──▶ stella ──▶ provider (your key)
             │
             ├─ tools: read · edit · bash · grep · ci · verify · …
             ├─ MCP servers (.stella/mcp.toml)
             ├─ memory & code graph (.stella/)
             └─ judge (goal mode) ──▶ "is the goal actually met?"

Stella sends your prompt to the model you selected, lets the model call tools to read and change your workspace, feeds the results back, and repeats until the task is done. In goal mode a second model judges the result from evidence, so the loop ends on a verified outcome rather than the worker's own say-so.

The documentation

• Getting Started — install, initialize a repo, configure provider keys.
• Inference Pipeline — triage → plan → witness → execute → verify → judge: how work gets proven, not just done.
• Context Engine — memories, the tree-sitter code graph, reflections, and the prompt-cache-native recall that makes them cheap.
• Agent Modes — Command Deck chat, one-shot runs, judged goal mode, CI monitoring.
• Agent Fleets — parallel workers in isolated git worktrees.
• Agent Tools — the built-in toolbelt, skills, MCP servers, custom commands, hooks, and permissions.
• Configuration & Settings — settings.json, per-agent engine config, credentials, and the scope hierarchy.
• Telemetry & Usage Analytics — local-only metering, the dashboard, budgets, and the files-touched ledger.
• Donate — keep independent, local-first tooling alive.
• Commands — the full command and flag reference.