Software Developer

Building tools that accelerate development, one agent at a time.

I'm Pablo Castañeda, a senior developer with 10+ years of experience, a lover of good coffee and new technologies. Through this blog I document the open-source AI projects I'm building — always with the goal of creating tools that improve development processes within companies and bring devs closer to a more friendly, easy-to-adopt agentic environment. These systems and projects seek to positively impact Customer Lead Time and Local Lead Time.

pablo@dev — agent session
$ agent run --goal "cut lead time" generating opportunity analysis detecting potential improvements brainstorming generating solution testing solution releasing solution to the community $
10+years exp.
OSSopen source
lead time
// shipped

Projects

A curated collection of the things I've built — with context on what they solve, how they work, and why they matter.

Open Source · AI Tooling

free-code

The first agent designed and built for development — flexible and open source.

The first agent to natively ship tools that accelerate development and bring value to devs and companies every day. Designed to accelerate every stage of a project: it doesn't just focus on understanding the codebase, writing and testing code — it's also flexible for organizations, with the ability for every company to modify it to fit their needs.

  • Any model — Anthropic, Google Vertex, or open-source models like Qwen Coder, Llama, and DeepSeek via LM Studio, Ollama, or any OpenAI-compatible endpoint
  • Vision Mode — free-code is the first agent with augmented reality: it can understand what you say and respond back. It can help you understand your surroundings — for example, taking screenshots of what you're looking at and giving you solutions or advice, something no other agent is capable of today.
  • Code graph — as you already know, context is everything. With native code-graph, free-code can generate code graphs of the project and understand it far better, in less time and spending fewer tokens. This takes the agent to the next level.
  • Agent browser — navigates the web natively over CDP: browsing, searching, and fetching whatever the task needs
  • Subagents — parallel agent sessions, each with its own context and tools, working on independent subtasks
  • Local RAG — today an agent needs more information in a secure and reliable way; with its local RAG, free-code can manage private knowledge bases, letting the agent extend its knowledge without retraining a model.
  • Skills & MCP — reusable behaviors, slash commands, and any MCP server plugged in as new tools
  • Webhooks — free-code is the first open-source agent able to receive external events as webhooks, something unthinkable not long ago.
Getting started
TypeScript Node.js VS Code Extension LLM RAG MCP
Open Source · Automation

agent-webhook-bridge

Until today, Claude could easily communicate with the outside world through MCPs or by making curl calls. But it couldn't receive events from the outside — that has changed with agent-webhook-bridge (AWB).

AWB gives Claude (I'm working on adding Cursor, Codex, and Copilot) an inbound door: an external event — a GitHub push, a CI failure, a Flowise flow, a plain curl — wakes the agent, with no polling and no need for the agent to be running beforehand. All of it secure: nobody can trigger it without your secret.

  • HTTP-triggered — one POST spawns the agent headless, as a fresh session or resuming a previous conversation
  • Secure by default — nothing is exposed unless you decide to; every hook gets its own secret, without it the webhook won't fire
  • Configurable per hook--workdir, --prompt-template, --permission-mode, --trigger/--queue, --consumer, --visible: each endpoint defines how its agent runs
  • Event protocol — events fire the agent the moment they arrive, or persist in SQLite as a queue for later consumption
Demo
TypeScript Node.js Webhooks Event-driven AI Agents
Open Source · First AI Botnet

agent-mesh

A hub where operators publish AI agents and anyone can send them tasks from a simple web UI.

The first working piece of the AgentMesh vision: operators publish AI agents running on their own machines, and anyone can send them tasks from a web UI. The hub never runs agents itself — it routes jobs and collects results, turning idle agent capacity into a shareable service.

  • Complete task loop — submit → run → result, with agents spawned headless in sandboxed workdirs
  • Session chaining — jobs can resume previous Claude sessions, so conversations continue across tasks
  • Remote access — shareable through a cloudflared tunnel, with per-user API keys that expire, rate-limit, and cap usage
  • Zero dependencies — a single-process Node hub: no build step, nothing to install
Create an agent, send a job
TypeScript Node.js SQLite AI Agents Claude Code
Open Source · MCP / AI Tooling

mcp-flowise

An MCP server that connects AI agents directly to Flowise chatflows and pipelines.

Turns your Flowise chatflows into tools any MCP-compatible agent can call: design a pipeline in Flowise's no-code UI and it immediately becomes an agent capability — no glue code in between.

  • Auto-discovery — queries the Flowise API and registers each chatflow as a named MCP tool
  • Typed inputs — every tool gets its own input schema, so agents call flows like native actions
  • Drop-in — a standard MCP server; works with any agent framework that speaks MCP
Demo
TypeScript MCP Flowise LLM Pipelines AI Agents
Open Source · AI Orchestration

The Target Project

A local hub that runs multi-step Claude Code and free-code workflows as one continuous, self-checking conversation — with templates, a waiting status for human review, and Slack notifications.

The Target Project creates one dedicated agent per workflow: its steps run strictly in order, always resuming the same session, so the whole workflow reads as a single conversation that advances step by step — with each step able to judge its own result, wait for a human review, or retry on a timeout before moving on.

  • Chained sessions — every step resumes the previous session (Claude Code or free-code), so the whole workflow is one continuous conversation from start to finish
  • Templates — save a reusable, ordered step list (with criteria, retries, retry intervals and manual-review flags) to seed new workflows or append to existing ones
  • Waiting status & human review — flag a step for manual review; once its work is accepted it holds as waiting and pauses the whole workflow until you press Continue
  • Judge-driven retries with intervals — steps with an acceptance criterion self-evaluate, and are retried with feedback and a configurable interval until they pass or the retry budget runs out
  • Settings & notifications — turn on Slack DMs so the hub pings you when a step starts waiting and when a workflow finishes
  • Redesigned conversation panel — a context-window meter, turn and token totals, and an Open-conversation button that resumes the session in a terminal
TypeScript React Vite Node.js node:sqlite Claude Code free-code AI Agents
// who I am

About Me

I'm a fullstack developer, a lover of frontend and backend technologies like React, Next, microfrontends, microservices, Node, Express, Java, .NET, Python, and Docker. I'm currently focused on AI-powered tools designed for developers. I believe the best tools are the ones that are easy to use, but also easy to adopt and that bring value.

My work centers on combining large language models, developer experience, and automation. I build things I actually use every day, which means I care deeply about reliability, performance, and security.

I look to positively impact KPIs and OKRs. I firmly believe that simply adopting tools isn't enough — they have to generate a positive impact on the metrics that matter to the business, like Customer Lead Time and Local Lead Time.

When I'm not shipping code, I'm exploring new AI models, contributing to open source, and thinking about how software development itself is changing.

// highlights
  • Double-digit reductions in CLT and LLT
  • Creator of skills, MCPs, and agents
  • Role model within the company
  • Early AI adopter
  • Expert in Claude, Cursor, Codex, and Copilot
  • Creator of free-code
  • Metric-first approach

Frontend

JavaScript React Next.js Microfrontends Context API HTML/CSS

Backend & data

Node.js MongoDB GraphQL RESTful APIs

AI

Custom AI Agents MCP Token Optimization Prompt Architecture Agent Orchestration Looping Flowise

AI-assisted tools

Cursor Claude Code GitHub Copilot CLI

Testing & QA

Cypress React Testing Library Unit/Integration/E2E A/B Testing

Methodologies

Agile Code Reviews CI/CD Mentorship
// in progress

Work in Progress

Things I'm currently building or exploring. No promises, no ETAs — just honest signal about where my head is at.

Phase 2 in progress

AgentMesh

A decentralized network where anyone can expose idle AI agent capacity, running sandboxed in local Docker containers, and earn points redeemable for USDT/BTC for the jobs the network routes to them. Phase 1 — the single-operator hub with a complete task loop — is built and open source (see agent-mesh under Projects), and phase 2 has started: the hub is shareable through a cloudflared tunnel with per-user API keys and rate limiting; remote nodes and the ledger are next.

Read the paper → Docs →
Idea

Coming soon

More experiments in AI tooling and developer automation.