An agent orchestration framework for solo builders who ship real products with AI.
You're one person. You have qulib, notquality.com, maybe a client project next month. You're using Cursor (or similar) to write code with an AI agent — and it's fast, until it isn't. The agent drifts. It pushes to main. It publishes a broken package. It forgets what you were working on yesterday. You end up babysitting the thing you hired to stop babysitting.
Tap-AgentOrchestrator fixes that. It's the operating system between you and your agents — a set of rules, skills, hooks, and state files that make your AI agent behave like a junior engineer with guardrails, not a rogue autocomplete.
Most AI dev tooling falls into two camps:
- "Let the agent do everything" — full autonomy, zero oversight. Great until it force-pushes to main or publishes the wrong version.
- "Chat-as-IDE" — you drive every keystroke. Fast for small edits, exhausting for real projects.
Tap-AgentOrchestrator is neither. It's the missing middle for solo operators:
- Your agent gets autonomy within boundaries — it can code, commit, create PRs, but it physically cannot merge, publish, or deploy without passing a Guardian gate.
- You get persistent memory across sessions — clock in Monday morning and the agent knows what you were shipping Friday, what's blocked, and what's next.
- You get Slack on your phone — walk away from the laptop and the agent pings you when a PR is created, a build breaks, or it's stuck waiting for your input. Not noise — structured lifecycle events with severity levels, threading per initiative, and muting controls.
- You get one framework across every project — scaffold once, and qulib, notquality.com, and your next side project all run with the same rules, same commands, same Slack channels.
- You get one set of rules across every LLM — write your guardrails once in
standards/, sync to Cursor, Claude, Codex, Copilot, or whatever ships next. No copy-pasting, no drift between platforms.
The target user is a solo curator like me: someone managing multiple real projects (not toy demos), who needs agentic help spun up fast and kept honest. You don't have a team to catch mistakes. The framework catches them instead.
| Concern | What's included |
|---|---|
| Roles | Conductor (planner), Orchestrator (sequencer), Guardian (gatekeeper), Model Router (tier-based model selection), Research / QA Sweep (project-specific helpers) |
| Control surface | 15 slash commands: /status, /start, /clock, /scope, /stop, /publish, /risk, /tradeoff, /adr, /checkpoint, /postmortem, /model, /notify, /portfolio, /sync |
| Persistent state | initiatives.json, sessions.json, gates.json — your work-in-flight ledger, survives Cursor sessions |
| Guardrails | Hard rules, autonomy levels, 7 gate checkpoints, mechanical hooks blocking dangerous actions |
| Model routing | Tier-based selection (reasoning / workhorse / fast / volume / vision) — right model for each task, swap providers in one config block |
| Slack integration | Lifecycle events (PR created, build failed, publish done, agent blocked) push to your phone via incoming webhooks. Per-project channels + a dashboard channel for cross-project visibility. Thread-per-initiative keeps channels clean. Mute by severity or event type. Walk away from the laptop, stay informed. |
| Multi-project | Portfolio registry at ~/.tapagent/. Per-project Slack channels + #tapagent-dashboard. /portfolio for cross-project status, blockers, daily digest. Consulting-ready. |
| Multi-LLM sync | Canonical rules in standards/ — sync to Cursor (.mdc), Claude (CLAUDE.md), Codex (AGENTS.md), Copilot (.github/copilot-instructions.md). One source of truth, every platform covered. |
| Scalability | Same framework across every project. One init.sh command per project. Works for side projects, works for clients. |
- Solo builders shipping real products. You're one person with multiple projects and you need your agent to act like a reliable junior, not a loose cannon. You want to delegate without losing the audit trail.
- Multi-project operators. You're running 3+ projects and you want them to feel like one consistent shop — same commands, same Slack channels, same guardrails.
- People who step away from the laptop. You start a task, go make coffee, and come back to a Slack thread that tells you exactly where things stand. Not a silent agent that may or may not have broken something.
- Engineers who care about rigor. ADRs, risk analysis, postmortems, and architecture decision records are first-class citizens — not afterthoughts bolted on after a production incident.
This is not for: large teams (use real PM tooling), throwaway scripts (overkill), or anyone who wants the agent to do everything autonomously (this framework intentionally adds friction at risk points).
# In your project root:
bash <(curl -s https://raw.githubusercontent.com/TapeshN/Tap-Agent/main/bin/init.sh)
# Or, if you've cloned Tap-Agent locally:
../Tap-Agent/bin/init.shThe scaffolder will:
- Copy
.cursor/rules/,.cursor/skills/,.cursor/hooks/into your project - Generate a
CLAUDE.mdfrom the template, customized to your project - Generate an empty
roadmap.json - Generate
.cursor/state/(gitignored) for live tracking - Append the right entries to your
.gitignore - Ask 5 questions (project name, type, build cmd, test cmd, release target) and write
project.config.json
After that, open the project in Cursor and say "what's next?". Your agent is now wearing the framework.
| Type | Examples | What changes |
|---|---|---|
npm-library |
qulib, CLI tools published to npm | Adds publish workflow, version-bump release PRs, registry verification |
web-app |
notquality.com, SaaS prototypes | Deploy-on-merge workflow, no publish, env var checkpoints |
cli-tool |
Single-binary CLI | npm or Homebrew publish, executable bit checks |
mobile-app |
Expo / React Native | EAS Submit workflow, build profiles |
internal-script |
One-offs, automations | Minimal — just clock and status |
research |
Data analysis, experiments | Artifact tracking, reproducibility checks |
Each type ships with a matching project.config.json in examples/.
PHILOSOPHY.md— design principles and what this framework rejectsdocs/01-architecture.md— layered model, separation of concernsdocs/02-agent-roles.md— Conductor, Orchestrator, Guardian, helpersdocs/03-state-model.md—initiatives.json,sessions.json,gates.jsondocs/04-guardian-gates.md— every gate, what it checks, when it pausesdocs/guardrails.md— behavioral guardrails (including commit discipline)docs/branch-strategy.md— branch naming, one-initiative-per-branch, ORCH-oriented flowdocs/context-ledger.md— audit trail for what context was loaded and whydocs/context-budgeting.md— per-class default context tiers and resetsAGENTS.md— global Cursor operating rules (Ask/Plan/Agent/Debug, class boundaries)docs/agent-classes.md— index of class docs (Conductor, Builder, Reporting, Librarian, Accountant)docs/agent-naming.md— Cursor UI vs framework roles (ORCH-012)docs/agent-lifecycle.md— session start/end, handoffs, one writer per branchdocs/report-stream-digest.md— ORCH-007 ReportStreamReporter pipelinedocs/slack-routing.md— ORCH-008 messageType → channel routingdocs/subagent-policy.md— when subagents may be proposed (not created by default)docs/deterministic-tooling.md— ORCH-010 planned tools catalogdocs/cost-accounting.md— cost source hierarchy (Cursor export → estimate → unknown)docs/commit-discipline.md— logical commits, checkpoints, Conventional Commitsdocs/05-slash-commands.md— full referencedocs/06-use-cases.md— per-project-type guidancedocs/07-consuming.md— how a project plugs indocs/08-extending.md— add a custom rule, skill, or gatedocs/09-risks-tradeoffs.md— known limits of this frameworkdocs/10-model-routing.md— model selection by tierdocs/11-notifications.md— Slack integration, event routing, setupdocs/slack-control-surface.md— Slack as Conductor decision channel (contracts)docs/slack-integration-roadmap.md— Phased Slack delivery roadmapdocs/12-multi-project.md— portfolio orchestration, multi-channel routingdocs/13-multi-llm.md— canonical standards, multi-LLM sync, platform targetsadrs/— architecture decision records
v0.7.0 — actively dogfooding via qulib and notquality.com. Built by Tapesh Nagarwal.