ApproachSolutionsTeamAI BlogContact
Talk to us →
PrivacyTermsLegal

Copyright © 2026 Baobab Tech. All rights reserved.

ApproachSolutionsTeamAI BlogContact
Talk to us →
Idea, AI Tools and Technologies  |  January 24, 2026

/help - best of mcp, skills & bash?

Skills and Knowledge for the network

Surely someone has thought of this… if so please let me know .. otherwise here goes:

Working names: /help, Clive (CLI Virtual Endpoints), Recline (Remote CLI)

Remote capabilities that feel like bash, use progressive disclosure like Skills, and work over HTTP. Any agent that can curl a URL can discover and use a /help endpoint.


The Idea

Skills are folders with markdown files. An agent reads what it needs.

/help is the same thing, but remote. A URL with endpoints. An agent fetches what it needs.

/help is for both services: do things, and for knowledge: informs things. A /help endpoint can be:

  • A service (process this PDF, analyze this data, do some deep research)
  • A knowledge base (here’s everything about our API, our policies, our products)
  • A reference (here’s the schema, the docs, the specs)
💡

My thought is having the simplest starting point: paste a URL into any agent or chatbot that can run an HTTP GET or curl command, which I think is the biggest simplification.. no connectors… let the agent web scrapers discover your information and services ahead of time or on the fly as the user pasted a simple url..


Agentic tooling history

MCP (November 2024)

MCP launched as an open standard for connecting AI to tools. Problem: tool definitions consume excessive tokens when you connect many servers.

Agent Skills (October 2025)

Anthropic introduced Skills as folders of markdown + scripts. Progressive disclosure: load metadata first, full instructions when triggered, supporting files when referenced.

AGENTS.md (2025)

OpenAI introduced AGENTS.md as “a README for agents.” Project-specific guidance in a predictable place.

Bash-Native Agents (2025)

Tools like Claude Code, OpenCode, and Vercel’s just-bash showed that bash is becoming the universal agent interface. Models understand --help. They navigate filesystems naturally.

The Gap

All of these work locally. What about easy remote knowledge and capabilities without the need for MCP connector setup?


/help: The Pattern

Every /help endpoint works like a CLI tool:

GET /help → what is this, what's available GET /topic/help → tell me more about this topic GET /topic/subtopic → go deeper

Responses are markdown. Human-readable. Agent-parseable.


Two Modes

Mode 1: Knowledge

A /help endpoint as a knowledge base:

GET /help
# Product Documentation Everything about our platform. ## Topics - `/api/help` - API reference - `/guides/help` - How-to guides - `/policies/help` - Terms, privacy, compliance - `/changelog/help` - What's new ## Search POST /search {"q": "authentication"}

An agent exploring this doesn’t need the full docs upfront. It fetches /help, sees the topics, fetches what’s relevant to the task.

Mode 2: Capabilities

A /help endpoint as a service:

GET /help
# PDF Tools Work with PDF documents. ## Available - `/extract/help` - Pull text from PDFs - `/fill/help` - Fill form fields - `/merge/help` - Combine documents ## Quick Example POST /extract {"url": "https://..."}

Same pattern. Discover first, use second.


Why This Works

For knowledge: Instead of dumping a whole knowledge base into context, agents can navigate it. Fetch the table of contents, drill into relevant sections, search when needed.

For capabilities: Instead of loading all tool schemas upfront, agents discover what’s available and how to use it on demand.

For both: Progressive disclosure. The agent only loads what it needs for the current task.


The Parallel

Skills (Local)/help (Remote)
A folderA URL
[object Object] at root[object Object] at root
Reference filesSub-endpoints
Scripts that runEndpoints that act
Agent reads filesAgent fetches endpoints

Implementation

Minimal:

from fastapi import FastAPI from fastapi.responses import PlainTextResponse app = FastAPI() @app.get("/help", response_class=PlainTextResponse) def help(): return """ # My Knowledge Base Information about X. ## Topics - `/overview/help` - What is X - `/setup/help` - Getting started - `/reference/help` - Full reference ## Search POST /search {"q": "your query"} """ @app.get("/{topic}/help", response_class=PlainTextResponse) def topic_help(topic: str): return load_markdown(f"{topic}.md") @app.post("/search") def search(body: dict): results = search_knowledge(body["q"]) return {"results": results}

That’s it. Serve markdown at /help. Add depth as needed.


Design Notes

On structure: There’s no required structure beyond /help at the root. Organize however makes sense for your knowledge or capabilities.

On search: Optional but useful for larger knowledge bases. A simple POST /search that returns relevant topic paths.

On versioning: Not really a thing. The /help describes what’s current. If it changes, the /help changes.

On registries: The /help endpoint is the registry. It lists what’s available.

On composition: One /help can link to others. It’s just URLs.


Use Cases

Documentation as /help: Your docs site, but agent-navigable. Agents fetch sections as needed instead of getting the whole thing.

Internal knowledge bases: Company policies, product specs, procedures. Agents can look things up.

API references: Instead of OpenAPI schemas, markdown descriptions agents can read and understand.

Service wrappers: Put /help in front of any capability. Describe it in markdown. Let agents discover how to use it.

Multi-agent retrieval: Complex RAG setups where agents navigate structured knowledge rather than vector-searching everything.


Comparison

AspectMCPSkills/help
TransportJSON-RPCFilesystemHTTP
DiscoveryUpfront schemaMetadata in prompt[object Object] endpoint
Context costHighLowMinimal
Knowledge or capabilitiesCapabilitiesBothBoth
Human readableNoYesYes

Prior Art

  • MCP Tool Search - Anthropic’s approach to reducing MCP context bloat
  • Agent Skills - The local pattern /help extends to remote
  • AGENTS.md - Similar spirit for project guidance
  • CLI-style MCP experiment - Earlier exploration of -help patterns

Summary

Skills showed that folders + markdown is enough for local agent capabilities.

/help shows that URLs + markdown is enough for remote knowledge and capabilities.

Any agent that can fetch a URL can use a /help endpoint. No SDKs. No schemas. No protocols. Just HTTP and markdown.


References

  • Anthropic. “Introducing the Model Context Protocol.” November 2024.
  • Anthropic. “Code execution with MCP.”
  • Anthropic. “Equipping agents for the real world with Agent Skills.” October 2025.
  • AGENTS.md
  • Vercel. just-bash
  • Linux Foundation. “Agentic AI Foundation.” December 2025.

Draft v0.2 - January 2026

PrivacyTermsLegal

Copyright © 2026 Baobab Tech. All rights reserved.