The Toolchain¶
23 tools that make the flywheel turn. Three are load-bearing; the rest are accelerators.
The core trio works together: Beads decomposes work into a dependency graph, BV uses that graph to tell each agent what to do next, and Agent Mail lets agents coordinate without blocking each other. Everything else plugs into this loop.
Core Tools¶
| Tool | What it does | Why it matters |
|---|---|---|
| Beads (bd/br) | Dependency-aware task graph: epics, tasks, subtasks with blocking relationships | The execution surface for all agent work. Without beads, agents randomly pick tasks or over-communicate. |
| BV (Beads Viewer) | DAG analysis, PageRank prioritization, --robot-next for automated assignment |
Answers "what should I work on next?" using graph theory instead of guesswork. |
| Agent Mail | Async inter-agent messaging with targeted recipients, file reservations, semi-persistent identity | Coordination without blocking. No broadcast-all, no ringleaders, no single points of failure. |
| NTM (Named Tmux Manager) | Agent cockpit: spawn, monitor, and manage agent sessions across 80+ commands | The human's control plane for tending 9+ agents across multiple projects. |
| UBS (Ultimate Bug Scanner) | 1000+ detection rules via AST-grep, 8 languages, deeper than standard linting | Catches semantic issues that linters miss: dead code paths, type mismatches, suspicious patterns. |
| CASS (Coding Agent Session Search) | Tantivy-powered search across 11 agent formats, sub-60ms results | Find what any agent did in any session. Essential for debugging cross-agent issues. |
| SLB (Simultaneous Launch Button) | Nuclear-launch-style confirmation with 4 risk tiers | Prevents accidental mass operations (killing all agents, wiping all state). |
| DCG (Destructive Command Guard) | SIMD-accelerated pattern matching, 50+ safety rule packs | Catches dangerous commands (rm -rf /, git push --force main) before they execute. |
Supporting Tools¶
| Tool | What it does |
|---|---|
| CM (CASS Memory) | Persistent memory layer for agents: short-term, episodic, and semantic recall across sessions |
| CAAM (Account Manager) | Sub-100ms auth switching between multiple AI provider accounts |
| RU (Repo Updater) | Multi-repo sync with AI-generated commit messages |
| MS (Meta Skill) | Skill management with adaptive selection (recommends which skills to use based on context) |
| RCH (Remote Compilation Helper) | Transparent build offloading to remote workers for heavy compiles |
| WA (WezTerm Automata) | Terminal hypervisor: automates multi-pane agent workflows |
| JFP (JeffreysPrompts CLI) | Browse and install prompts as Claude Code slash commands |
| SRPS (System Resource Protection) | Auto-deprioritize background processes to protect agent sessions |
| APR (Automated Plan Reviser) | Iterative spec refinement with extended reasoning |
| PT (Process Triage) | Bayesian detection of zombie and runaway processes |
| XF (X Archive Search) | BM25 + semantic search over Twitter/X archives |
| TRU (TOON Rust) | Token-optimized notation for LLM context compression |
| MDWB (Markdown Web Browser) | Convert websites to clean markdown for LLM consumption |
| S2P (Source to Prompt TUI) | Interactive code-to-prompt generator with file selection |
| CAUT (Usage Tracker) | Track LLM provider usage and costs across accounts |
Setup¶
Transform a fresh Ubuntu VPS into a complete agentic environment:
curl -fsSL "https://raw.githubusercontent.com/Dicklesworthstone/agentic_coding_flywheel_setup/main/install.sh?$(date +%s)" | bash -s -- --yes --mode vibe
Idempotent. If interrupted, re-running resumes from last completed phase.
Monthly cost: VPS (\(40-56) + Claude Max (\)200) + ChatGPT Pro (\(200) = ~\)440-656.