Poncho, a general agent harness built for the web

Name: Poncho, a general agent harness built for the web
Availability: InStock
Author: heycesr

by heycesr·Feb 18, 2026·4 points·2 comments

Visit Project View on HN

AI Analysis

●●SolidNiche GemShip ItBig Brain

Git-native agent framework with local dev parity to production deployment.

Strengths

•Git-first approach to agent config and skills enables reviewable diffs and rollbacks.
•Serverless-friendly by design; same AGENT.md works locally and in production.
•Agent Skills open standard portability across Copilot, Cursor, VS Code reduces lock-in.

Weaknesses

•Early-stage (beta, breaking changes expected); only .md and JS/TS skills, Python/bash limited.
•Crowded space: Claude Code, OpenClaw, Cursor, Continue already solve agent-building for many teams.

Post Description

Hi HN! I'm building poncho, a framework for building custom AI agents that are version-controlled in git, developed locally, and deployed as isolated endpoints (serverless-friendly by default).

Poncho agents follow the same conventions as openclaw or claude code so they probably feel familiar.

You can talk to them via terminal or web ui to build new skills or configure the agent, and they're compatible with the Agent Skills open spec, so you can port your skills. Small caveat: right now it's compatible with .md skills and js/ts scripts, but many skills are bash or python-based. I'm still figuring out what would be the best way to extend support for those skills without overcomplicating it, if anybody has any ideas I appreciate it!

Some features:

- Git-native: agent behavior, skills, and tests live in your repository (reviewable diffs + easy rollbacks). - Single-file agent definition: define runtime config + instructions in AGENT.md (YAML frontmatter + prompt content). - Skills you can ship: AgentSkills-style skills/*/SKILL.md plus TypeScript/JavaScript scripts under scripts/. - MCP support: connect remote tool servers and inject required environment variables through config. - Conversation-first API + streaming: stored conversations with SSE streaming responses and tool events. - Pluggable storage + memory: local files for dev or hosted stores (e.g. Upstash), with optional persistent memory + recall. - Testing + observability: poncho test workflows and OpenTelemetry traces/events.

Now the cool thing about poncho agents is that they're super easy to deploy to Vercel/Fly/Lambda/etc, so you can share them with anybody in your team and enable non-technical people with custom skills.

I also built and deployed a couple example agents here: - https://github.com/cesr/product-agent - https://github.com/cesr/marketing-agent

I'd love some feedback, I started building poncho because I wanted a fast and easy way to build and share agents with my team. Let me know what you think!