ADDA SDLC is a structured agentic framework: an AI coding agent works inside a hardened, ephemeral container and persists all development state through GitHub.
ADDA runs an AI coding agent inside an isolated, ephemeral container. The agent operates with no persistent local state — all work is committed and pushed to GitHub. Network access is enforced by a per-session Envoy proxy sidecar on the host, with a default-deny domain allow-list.
The runtime is a tier stack:
- Tier 1 (
adda-dev-runtime) — generic, AI-tool-agnostic base image. Provides the hardened container infrastructure: entrypoint, proxy bridge, system tools, and Bun scripting runtime. - Tier 2 (
proto-adda) — AI harness layer. BuildsFROMTier 1, adds Claude Code, and delivers the SDLC configuration. - Tier 3 — the project being developed. A standard GitHub repo that uses the runtime as its development environment.
This repository ships Tier 1 and proto-adda (Tier 2 for Claude Code). See docs/adda-dev-runtime-design.md for the full conceptual design.
To launch a session on the host, use the adda-dev-launcher — it provides the launcher script, Envoy proxy configuration, and host setup instructions.
| Document | Contents |
|---|---|
docs/adda-dev-runtime-design.md |
Conceptual model: design goals, threat model, tier architecture |
docs/adda-dev-runtime-technical-design.md |
Implementation reference: entrypoint sequence, authentication, networking, image build |
docs/proto-adda.md |
proto-adda (Claude Code harness) specifics |
docs/self-hosting.md |
Developing this repo from within itself |
- Linux only. Ubuntu 24.04 is the tested platform. No plans to support other platforms.
- Ephemeral by design. The container has no persistent storage. Push commits to GitHub before ending a session — anything not pushed is lost.
- Personal project. Features and design reflect personal workflows. PRs from external contributors are not being accepted. Feel free to fork and adapt.
MIT — see LICENSE for details.