Skip to content
PRODUCT / 2026

FDE Agent Starter

Batteries-included template for shipping a forward-deployed AI agent — evals in CI, Langfuse traces, one-command Fly.io deploy, and a bundle of Agent Skills.

View on GitHub
STATUSActive
YEAR2026
AREAS
  • Open Source
  • AI Agents
  • Vercel AI SDK
  • Langfuse
  • Agent Skills
  • TypeScript
THECODERPANDA / FDE AGENT STARTER
THE PROJECT / 01

FDE Agent Starter

← Back to the index

The Problem

Every forward-deployed AI engagement rebuilds the same scaffolding.

Day 1: chat UI. Day 3: someone asks "how do we know it won't get worse tomorrow?" and the real work starts — evals, traces, a deploy pipeline, a way to add tools without regressing the ones you already have, a way to A/B two prompts without arguing about vibes.

I have rebuilt that scaffolding at least four times. So I packaged it.

What It Is

An opinionated open-source template for shipping a production AI agent, not a demo.

Layer What ships
UI Next.js App Router + Vercel AI SDK chat (useChat + streaming /api/chat)
Tools Typed with Zod: web_search, sql_query, read_file (patterns, not just utilities)
Evals JSONL golden cases + typed assertions + CI gate on every PR + nightly cron
Traces Langfuse wrapper — silent no-op without keys, full span per tool call with keys
Deploy Multi-stage Dockerfile (non-root, ~140 MB) + fly.toml with /api/health
Providers OpenAI · Anthropic · Gemini · DeepSeek · OpenRouter, auto-detected, env-driven model IDs
Skills 5 Agent-Skills runbooks: agent-eval, prompt-diff, trace-debug, deploy-agent, add-tool

Quickstart

git clone https://github.com/thecoderpanda/fde-starter-kit fde-agent
cd fde-agent
npm install
cp .env.example .env    # add one provider key
npm run dev             # http://localhost:3000
npm run eval            # run the golden set

Five commands to a working agent with evals, traces (optional), and a deploy path.

Principles

  • Opinionated defaults, permissive escape hatches — one framework, one deploy target, one eval format. Rip out any piece as you outgrow it.
  • The scaffolding is the product — evals, traces, CI, and deploy are shipped by default, not homework.
  • Multi-provider from day one — three-tier model precedence (AGENT_MODEL → AGENT_MODEL_<PROVIDER> → fallback) means swapping models is an env change.
  • Skills over docs — five self-contained runbooks encode the FDE workflow so your coding agent can do the FDE tasks with you.
  • MIT, forever — clone it, fork it, ship it, sell what you build on top.

Roadmap

  • evals — parallel runner with a rate-limited queue
  • agent — LangGraph-style branching for multi-step workflows
  • skills — cost-report skill (per-provider spend from eval runs)
  • deploy — Cloud Run + Render config alongside Fly.io
  • observability — OpenTelemetry exporter next to Langfuse

Status

Live. Public on GitHub, CI green, five skills shipped in the ./skills/ bundle. Actively iterating on the eval harness and skill set — issues and PRs welcome, especially new skills.

Licensed under MIT. Star the repo to follow along.