Visual AI workflow automation with MCP
AgentLoom
A visual builder for AI workflows with durable execution, multi-model LLMs and native Model Context Protocol tools.
My role: Design and full-stack development: canvas, runtime, AI layer, security and deployment
- Agents
- MCP
- Nuxt 4
- Inngest
- Vercel AI SDK

Overview
AgentLoom bridges static no-code automation (Zapier, Make) and autonomous AI agents. Teams visually chain LLMs, APIs, branching logic and tool-calling agents into durable, fault-tolerant execution graphs — then schedule them or trigger them by webhook.
The challenge
No-code tools like Zapier and Make are good at fixed sequences, but they struggle when a step needs judgement: classifying a lead, extracting fields from an email, or letting an agent decide which tool to call. Custom agent code, on the other hand, is hard for non-developers to change.
AgentLoom set out to combine both: a visual canvas anyone can edit, with AI nodes and autonomous agents that run reliably in production, can use external tools through the Model Context Protocol, and never leak the API keys they depend on.
How it works
- 1Webhook / Cron
- 2DAG validation
- 3Inngest step
- 4LLM / Agent
- 5MCP tools
- 6Result + cost
Architecture
The editor is a Nuxt 4 and Vue 3 application with a Vue Flow canvas offering 14 node types: triggers (webhook, cron, manual), AI nodes (text, structured JSON extraction, zero-shot classification, summarisation), tool-calling agents, logic (if/else, delay, transform) and integrations (HTTP, Resend email). Before a run starts, the client checks the graph for cycles and validates it as a DAG.
Execution runs on Inngest. Each node becomes a durable step, so a failure retries with exponential backoff and resumes from that node rather than restarting the workflow, with no serverless timeout. AI nodes use the Vercel AI SDK to switch between Gemini, OpenAI and Anthropic models; agent nodes run a multi-step loop with automatic tool invocation and a configurable step limit.
MCP support is full JSON-RPC 2.0 against remote MCP servers with automatic tool-schema reflection, protected by SSRF guardrails that block private-subnet requests. Credentials are encrypted with AES-256-GCM through Web Crypto. Webhooks are rate-limited with an Upstash Redis token bucket (30 requests per minute) and payload-size checks, and sessions use Better Auth with CSRF protection. Neon PostgreSQL with Drizzle stores workflows and run history.
What I built
Node canvas
Vue Flow designer with 14 node types: triggers, AI (text, JSON extraction, classification, summarisation), agents, logic and integrations. Cycle detection and DAG validation before every run.
Native MCP
JSON-RPC 2.0 integration with remote MCP servers, automatic tool-schema reflection, and SSRF guardrails that block private-subnet scanning.
Durable execution
Inngest gives step-level state recovery, retries with exponential backoff and no serverless timeouts.
Cost visibility
A live execution drawer shows per-node input/output, token counts and estimated USD cost.
Credential vault
AES-256-GCM authenticated encryption for API keys and tokens; secrets are never logged or shown in plaintext.
Defensive networking
Upstash Redis token-bucket rate limiting (30 req/min per webhook), payload size limits, Better Auth sessions and CSRF protection.
Engineering decisions
- Validate the graph on the client before a run so broken workflows fail in the editor, not in production.
- Run each node as a durable Inngest step so a retry resumes from the failed node instead of restarting the workflow.
- Show cost per node so teams can move simple classification to cheaper models.
Results
- Live demo at agentloom-zeta.vercel.app with a demo workspace.
- 14 node types, including autonomous agents and native MCP tools.
- Per-node input/output inspection, token counts and estimated USD cost for every run.
- Serverless deployment on Vercel, Neon and Inngest with no servers to manage.
Stack
- Frontend
- Nuxt 4 (Vue 3, TypeScript), Tailwind CSS v4, Vue Flow, Pinia
- Runtime
- Nitro / H3 on Vercel, Node.js 20+
- AI
- Vercel AI SDK — Gemini, OpenAI, Anthropic
- Orchestration
- Inngest (durable execution, scheduling, event bus)
- Data
- Neon PostgreSQL, Drizzle ORM, Upstash Redis
- Security & QA
- Better Auth, Web Crypto AES-256-GCM, Vitest, Playwright
Need something like this?
I can build a version of this for your product, your data and your stack.
Start a project