Archived. When Anthropic and OpenAI shipped their own CLI agents, this project became unnecessary. But building it is what taught me where the industry was going — and that led to Terrarium.
CLI agent for software development, built in Rust. Started January 2023.
Back then, AI-assisted coding meant autocomplete in your IDE. Copilot suggested lines. Cursor was just getting started. Nobody was building autonomous CLI agents that could read your project docs, orchestrate multi-step tasks, and work across LLM providers.
That's what Companion was. A bet that the terminal agent — not the IDE plugin — was the right shape for AI-assisted development.
- Autonomous coding agent with 19 commands: generate, review, test, translate, commit, PR descriptions, changelogs, diagrams
- Knowledge-aware: pointed it at your docs, codebase, GitHub repos, Notion pages, URLs, YouTube transcripts, or shell output, and it used that context for every task
- Graph-based task orchestration: multi-step workflows defined in YAML, executed as dependency graphs (petgraph, topological sort)
- Multi-model: OpenAI, Claude, Gemini, Ollama — with parallel map-reduce across instances
- Extensible: custom commands and reusable task plans via YAML
Claude Code launched. Then Codex CLI. The industry validated the bet: autonomous terminal agents with project awareness are how AI-assisted development actually works. Not autocomplete.
Companion became redundant overnight. But the year and a half I spent building it gave me something the tools themselves don't: a deep understanding of how agents should work, what breaks at scale, and what's missing from all of them.
That's what led to Terrarium — the cloud platform for agent workflows.
Rust, tokio, clap, serde, petgraph, octocrab, reqwest, ollama-rs, tiktoken-rs. Clean architecture (domain/infrastructure/application). Full docs in docs/.
MIT