byoa v0.1.0 — Bring Your Own Agent

byoa v0.1.0 — Bring Your Own Agent

First public release. A local-first coding agent harness: point it at any model, pick a prompt
tier sized to that model, and let it run shell + file tools inside a bubblewrap sandbox whose
policy you control.

Install (use a release artifact, not the source tree)

Standalone (needs Node >= 26; no npm install):

tar -xzf byoa-0.1.0-linux-x64.tar.gz
install -m755 byoa ~/.local/bin/byoa
byoa --version

npm-installable tarball:

npm install -g ./byoa-0.1.0-npm.tgz
byoa --version

Both contain the bundled JavaScript build. The TypeScript source is not a runnable artifact:
Node does not type-strip files inside node_modules, so releasing the source tree would not
install.

Included

• Streaming agent loop with tool calls, cancellation and approvals.
• Providers: OpenAI-compatible (including LM Studio, Ollama, llama.cpp, vLLM) and Anthropic.
• Session persistence with resume and usage counters.
• Tools: shell, fs_read, fs_write, fs_edit, fs_patch.
• Sandbox: policy → bubblewrap argv, read-allowlist skeleton root (no --ro-bind / /), network
  denied by default, cleared environment, ulimit limits, fail-closed strict mode.
• Three prompt tiers (lite / normal / large) owning persona, context budget, per-result
  truncation and output cap.
• Obsidian-backed memory: markdown notes, SQLite FTS5 index, memory_recall / memory_remember,
  namespaced to memory/byoa/ so the harness only reads what it wrote.

Not included yet

TUI, Web and Electron adapters, MCP, subagents, web fetch/search.

Requirements

• Node.js >= 26
• Linux with bubblewrap for sandboxing. Without it, strict (the default) refuses to run
  rather than silently degrading.

Quality

184 tests passing, including real-bubblewrap isolation tests; tsc --noEmit clean.

---

AI-generated with OpenCode — primary model: DeepSeek V4.1 Flash.