Braid

A toolkit for Claude Managed Agents.

Warning

Experimental — not for production use. Braid is an early-stage research project. APIs, flows, and outputs are unstable and may change or break without notice. It is provided as-is, without warranty of any kind, for experimentation and demonstration only. Do not rely on it for production workloads, sensitive data, or anything you cannot afford to lose.

Six tools

• Agent — the job description.
• Environment — the private office.
• Session — the workday that survives sleep.
• Skills — table of contents, not textbook.
• Vaults — agent knows the lock; session brings the key.
• Outcomes — the grader.

Setup

Requires Claude Code and bun.

export ANTHROPIC_API_KEY=sk-ant-...
export FAL_API_KEY=...          # for video / image generation (vault-injected)
export VERCEL_TOKEN=...         # consumed host-side by post-session hooks; never enters agent session

Secrets handling follows NIST SP 800-204C build/deploy separation and Anthropic's vault docs. FAL_API_KEY is injected by the Fal MCP proxy and never enters the sandbox. VERCEL_TOKEN stays on the host and is consumed by the post-session hook after the agent's session ends. See SECURITY.md for the threat model and disclosure policy.

Use

Inside Claude Code:

/braid                  # interactive
/braid list
/braid setup <flow>
/braid run <flow>
/braid pull <flow>
/braid purge

Flows

Each lives in flows/<name>/ with a flow.yaml and agent YAML files.

• ad — director + producer + critic + sentinel.
• fundraiser — videographer (Fal MCP) + web-builder (Vercel) + director. Ships a live one-pager with a hero video. live
• final-inning — 3-shot fundraiser, little-league field. live
• homecoming — 3-shot fundraiser, veteran service dogs. live
• quiet-rebellion — 3-shot fundraiser, scrubland conservation. live
• pop-quiz — 3-shot deadpan storyboard (goose substitute teacher) with a mediabunny-composited final.mp4.

Reading order

• SKILL.md — design + canonical patterns (brief substitution sandbox, host-side post-session hooks, MCP host allowlist, composite recipes).
• SECURITY.md — disclosure policy and threat model.
• CONTRIBUTING.md — TDD workflow, PR strategy, authoritative-source citation requirement.
• flow.schema.json — the canonical shape of flows/<name>/flow.yaml.

Configuration highlights

Each flows/<name>/flow.yaml can declare:

environment:
  networking:                 # defaults to limited with empty allowed_hosts
    type: limited
    allowed_hosts: [api.example.com]
    allow_mcp_servers: true

run:
  sentinel_context_store: projectStore   # store key for the sentinel diagnostic agent
  post_session_hook:                     # host-side hook for credentialed work
    command: bun run "$BRAID_FLOW_DIR/../../.claude/skills/braid/post-hooks/vercel-deploy.ts"
    env_passthrough: [VERCEL_TOKEN]      # strict scope; nothing else reaches the hook
    timeout_ms: 300000