Skillget

claude-agent-team

mcp
Guvenlik Denetimi
Uyari
Health Uyari
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Low visibility — Only 5 GitHub stars
Code Gecti
  • Code scan — Scanned 12 files during light audit, no dangerous patterns found
Permissions Gecti
  • Permissions — No dangerous permissions requested

Bu listing icin henuz AI raporu yok.

SUMMARY

CAST: multi-agent orchestration framework for Claude Code with hook-enforced quality gates and observability.

README.md

CAST — Swarm control plane for Anthropic Agent Teams

CAST

BATS Tests
Version
Agents
Tests
License
Shell

CAST is a production control plane for Claude Code built on three pillars: hook enforcement (every agent change is gated by validators — cast-validate-all-hooks.sh runs in CI and hookSpecificOutput shape is contract-validated), audit trail (cast.db with 31 tables records every session, agent run, routing decision, quality gate, and memory write), and a typed agent registry (23 agents, model-assigned across haiku 4.5 / sonnet / opus tiers, quality-gated, with frontmatter contracts). Define a workflow once; specialist agents plan, implement, review, test, and commit — automatically.

CAST Framework

Installation

Homebrew

brew tap ek33450505/cast && brew install cast

Manual install

git clone https://github.com/ek33450505/claude-agent-team.git
cd claude-agent-team
bash install.sh

Quick Start

docs/tutorial/getting-started.md — install, verify, and run cast status in 5 minutes.

What Makes CAST Different

  • Quality gates that actually enforce. Raw git commit and git push are hard-blocked by hooks. Code changes mandate a reviewer pass. You cannot skip this.
  • 23 specialist agents, pre-configured. Each has a bounded scope, a model tier, and a thinking budget. code-writer implements; code-reviewer reviews; commit commits. They don't cross lanes.
  • SQLite audit trail, fully local. Every agent dispatch, tool call, and token spend logs to cast.db on your machine. No SaaS dashboard, no cloud lock-in.
  • 1098 BATS test cases with 0 failures. Every hook script and utility is covered. CI runs on both macOS and Ubuntu on every push.

Documentation

Guide Description
Tutorial Install CAST and run your first agent dispatch
Compatibility Matrix Claude Code version requirements and known breakages
Hook Authoring Guide Write, test, and install custom hook scripts
Full Docs Index All documentation with one-line descriptions

Demo: cast status + orchestrate run — recording in progress

Table of Contents

What is CAST? · Architecture · Agents · Hooks · Observability · Ecosystem · Testing · Contributing · Stats

What is CAST?

CAST transforms Agent Teams into a production control plane:

  • Swarm composition in YAML. Define teams, assign roles, set quality gates. CAST bootstraps worktrees, seeds teammates with identity + prompts, manages peer messaging.
  • Structural quality gates. Code changes mandate a reviewer pass. Raw git commit and git push are hard-blocked by hooks.
  • Full observability. Every session, task, peer message, and token spend logs to cast.db (SQLite).
  • Local model routing. Haiku agents route to Ollama; cost per swarm drops 40-60% with LiteLLM proxy.

Your First Workflow

  1. Plan — /plan add user auth feature → planner writes an Agent Dispatch Manifest.
  2. Execute — /orchestrate next → code-writer implements, code-reviewer checks, test-runner verifies, commit agent stages.
  3. Ship — /ship → tests, CI sanity check, push, journal entry.

Architecture

Every CAST operation follows a four-stage flow: (1) hook validation — pre-tool hooks enforce quality gates and block non-compliant writes before any change lands; (2) agent dispatch — the typed agent registry routes work to the correct model tier (haiku 4.5 for review/commit, sonnet for implementation, opus for migration review); (3) memory injection — each agent receives relevant prior-session context from its ~/.claude/agent-memory-local/<name>/MEMORY.md on startup; (4) cast.db audit — every session, routing decision, quality gate result, and token spend is appended to the local SQLite database. See docs/architecture/ARCHITECTURE.md for the full guide including Agent Teams comparison table.

CAST swarm architecture

Personal Overlay

CAST ships in two layers: core (always installed) and personal (optional, --personal flag). New clones get a trustworthy, generic installation; rules-personal/ ships empty for clones to populate. See docs/personal-overlay.md.

Swarm System

CAST swarms are defined in YAML and bootstrapped with cast swarm bootstrap. Teams get isolated worktrees, agent identity, peer messaging, and quality gates. See docs/swarm.md.

Agent Constellation Dashboard

Constellation — force-directed graph showing agent nodes, task satellites, token heatmaps, peer messages, and hook audit trails in real time. Part of claude-code-dashboard. See docs/dashboard.md.

Cast Desktop — Tauri 2 native app surfacing the same observability layer with an embedded PTY terminal, command palette, and 11 dashboard views. See cast-desktop (v0.1.0).

Agents

23 core specialists. Each is a markdown file in ~/.claude/agents/ with YAML frontmatter defining model, memory, isolation, and thinking budget tier. Agent responses validate against JSON schemas in schemas/. See docs/agents/AGENT-ROSTER.md for the full table with model tiers and thinking budgets.

Key agents: code-writer, debugger, planner, researcher, security, code-reviewer, commit, push, test-writer, devops, bash-specialist, migration-reviewer, api-contract, dep-auditor, perf-sentinel.

Token Efficiency & Cost Optimization

Model tiering, response budgets, Ollama local routing, laconic mode, and RTK compression achieve ~30-50% token spend reduction. See docs/TOKEN-OPTIMIZATION.md.

Optional: Local-first cheap-mode (claude-code-router)

For local Haiku-tier work without API spend, install claude-code-router and run ccr instead of claude. CCR proxies all model calls to local Ollama (default: deepseek-coder:latest). Opt-in per session — vanilla claude continues to use Anthropic API.

Hooks

Hooks cover the full swarm lifecycle: SessionStart, TaskCreated, WorktreeCreate, PreToolUse:Bash (commit guard), PostToolUse, PostCompact, SessionEnd. See docs/hooks/authoring-guide.md.

Observability & cast.db

SQLite WAL mode at ~/.claude/cast.db — append-only, never truncated. Stores swarm_sessions, teammate_runs, agent_runs, routing_events, stream_events. See docs/observability/OBSERVABILITY.md.

Peer Messaging & Gossip Protocol

Teammates communicate via cast.db message bus — no central broker, fully decentralized. See docs/architecture/ARCHITECTURE.md.

Multi-Agent Pipelines

The /orchestrate skill executes Agent Dispatch Manifests (ADM) — JSON plan files defining sequential/parallel agent batches. owns_files prevents write conflicts between parallel agents. See docs/architecture/ARCHITECTURE.md.

Agent Memory & Persistence

Each agent accumulates domain knowledge in ~/.claude/agent-memory-local/<name>/MEMORY.md. Features: FTS5 full-text search, relevance scoring, shared pool, procedural memory, session distiller. See docs/architecture/ARCHITECTURE.md.

Routines: Scheduled Workflows

11 built-in time-triggered or event-triggered agent jobs — daily briefings, inbox triage, standup, weekly cost reports, and more. Manage with cast routines list / cast routines trigger <name>. Full guide: docs/routines.md.

Project Structure

agents/ · rules-{core,personal}/ · docs/ · schemas/ · scripts/ · swarm-configs/ · tests/ · .github/workflows/

Runtime installs to ~/.claude/ — agents, memory, plans, swarm sessions, cast.db, scripts.

Testing

115 BATS test files covering core hooks, swarm bootstrap, message bus, database migrations, guard logic, event emission, and memory persistence. 0 failures. BATS is installed via package manager — brew install bats-core (macOS) or apt-get install bats (Ubuntu). Run with bats tests/.

Version History

Full changelog: CHANGELOG.md.

CAST Ecosystem

CAST is one of 13 source repositories in a connected ecosystem — each solves a piece of the multi-agent workflow puzzle. All are open-source and actively maintained. See docs/ecosystem.md for the full repo table and install commands.

Tier Repos
Core Framework cast-hooks, cast-agents, cast-memory, cast-observe, cast-security, cast-parallel
Tooling cast-claudes_journal, cast-dash, cast-time, cast-routines, cast-doctor
Observability claude-code-dashboard, cast-desktop

New to CAST? Quick Start · Agent Roster · Architecture

Already using CAST? Changelog · Hook Authoring Guide · Full Docs

Used In / Built With CAST

Star History

Star History Chart

Contributing

Contributions are welcome — CAST is built in the open and actively developed. New agents, shell script fixes, BATS test coverage, and documentation improvements are all fair game.

Good first issues: good first issue label — curated entry points with clear scope and test expectations.

See CONTRIBUTING.md for the full workflow. Open an issue first for non-trivial changes.

Support & Portfolio

Star this repo if CAST is useful — visibility helps.

Built by Ed Kubiak as a showcase of production-grade multi-agent AI tooling. Portfolio →

License

MIT — see LICENSE. Built by Edward Kubiak — full-stack engineer, Claude Code expert. CAST Portfolio: castframework.dev

Stats

23 agents | 1098 test cases | 19 commands | 16 skills

Yorumlar (0)

Sonuc bulunamadi