STC — Software Team Collective

The Dev Agency OS for builders outgrowing single agents. Talk with a team of agents to shape goals, spec, and roadmap — like briefing a real company. Then the agency executes: roles, handoffs, verification gates, delivery pipeline. Local-first, MIT-licensed, built for 24/7 homelab operation.

5-minute quickstart

Prerequisites: Docker, a homelab box or VPS (Strix Halo / DGX Spark-class for 24/7), and one of: an API key, an OpenCode Go subscription, or an OpenCode Zen balance.

git clone https://github.com/Heretek-AI/STC && cd STC
cargo run -q -p studio-cli -- init --repo .   # writes .env (rootless default)
studio up                   # brings up the full compose stack
studio logs --follow        # watch the agency work

studio init refuses privileged-dev without typed confirmation and stamps whether the credential was live-pinged (wizard-written keys are marked NOT live-pinged — verify them in the cockpit Providers UI, which holds the live /models proof flow). Fallback: cp .env.example .env and edit by hand.

Open the cockpit (WebUI beta) at the printed address, or run studio-tui for the terminal.

Homelab hardening + 24/7 ops: see docs/homelab-quickstart.md (lane image build, UID/GID smoke test, volume drill, dev-mode guardrails).

The autonomy dial

New projects start at full authority: the LLM manager negotiates scope with you, dispatches into the DAG, and lands through the gates. Nothing needs a tap — the audit log is the safety net.

Prefer supervision? Turn the dial to advisory: scope decisions, dispatches, and merges pause for your approval (desktop, TUI, or phone via the approval relay). Per-project setting, remembered per repo.

studio autonomy --db studio.db --repo .                  # show current mode
studio autonomy --db studio.db --repo . --mode advisory  # pause for approval

Destructive ops (deploy, secret rotation, infra mutation, incident-resolve) always require explicit approval in both modes. The cockpit header and TUI status line show the live mode.

How work flows

scope → DAG → dispatch → review → merge, driven from the cockpit:

1. Scope — converse with the manager: goals, spec, roadmap.
2. Dispatch — zero-model scheduler assigns lanes (pi specialists for cost, opencode for heavy sessions, routed by measured $/task).
3. Review — evidence receipts (tests, syntax, fallow audit, Semgrep); tap-outs ("done" with no evidence) are blocked and sent back.
4. Merge — freeze → one bounded correction → burned acknowledgement → Bors lands. Nothing lands unverified.

Roles

v1 ships the full catalog (~30 roles: engineers, QA/security, creatives, management, ops) authored once and emitted to pi, opencode, and omp targets. Community roles live in the separate STC-Marketplace repo — core-team published, PR-reviewed.

Models

Bring your own keys, or point at OpenCode Go (flat-rate infinite-coding tiers) or Zen (pay-per-request) as presets. Preconfigured tiers per role (flagship / workhorse / disposable); override anything per role in the cockpit model console.

Docs

• AGENTS.md — working contract for agents (and humans) in this repo. Generated from code sources; never hand-edit derived sections.
• GOAL.md — standing operating goals for agents.
• plans/ — roadmaps, spike receipts, audits.

Status

v0.1 target: full docker stack + WebUI beta. See GOAL.md for operating goals and plans/ for the roadmap.