The Four-Layer Agentic Web
Your website is invisible to AI agents. They can't find it, can't understand it, can't use it, and can't talk to it. We fixed that.
AccountsOS implements a complete 4-layer architecture that makes our platform fully accessible to AI agents — from ChatGPT recommendations to autonomous agent-to-agent accounting workflows. Every layer is live and working. Here's how.
Most websites are invisible to AI
When someone asks ChatGPT “What's the best accounting software for a UK limited company?” — your product is only in the answer if AI agents can find, understand, and trust your content. Most can't.
Invisible to crawlers
AI bots blocked by robots.txt, no machine-readable files, no agent discovery endpoints.
Unstructured content
No answer capsules, no schema markup, no structured API. Agents can't extract what you do or who you serve.
No agent interface
No MCP server, no A2A endpoint, no way for agents to query your product or act on behalf of users.
Discovery — AI agents find us
Before an agent can recommend AccountsOS, it needs to know we exist. We publish machine-readable files at well-known paths that every major AI system checks.
/llms.txtLLM instruction file
Plain-text summary of what AccountsOS does, available tools, auth method, and quick-start curl examples. The first thing an AI reads.
/context.mdDeep context document
Full product context: ICP, pricing, features, API docs, UK tax specifics. What agents read for complete comprehension.
/mcp.jsonMCP manifest
Machine-readable manifest listing our MCP endpoint, protocol version, and all 13 available tools with their scopes.
/.well-known/agent.jsonAgentCard (A2A discovery)
Google A2A protocol. Describes Finn (our AI), its skills, authentication, and the endpoint other agents call.
Our robots.txt explicitly allows 16 AI crawlers — GPTBot, ClaudeBot, PerplexityBot, Google-Extended, and 12 more — with full access to agent endpoints.
Comprehension — AI agents understand us
Discovery tells agents we exist. Comprehension tells them what we do, who we serve, and how to talk about us accurately. This is how you control what ChatGPT says about your product.
/api/ai/contextStructured context API
Returns JSON: company info, features, pricing tiers, all 13 tools, MCP endpoint, UK tax context. Agents call this for machine-readable understanding.
Answer capsulesOn every content page
1-2 sentence direct answers at the top of every guide. AI models cite the first 30% of content 44% of the time — we put the answer there.
JSON-LD schemaOn every page
Article, FAQPage, HowTo, WebApplication, Organization, SoftwareApplication — structured data on all 400+ pages.
When ChatGPT or Perplexity answers 'What accounting software is best for UK Ltd companies?' — the answer capsule and structured data is what they cite.
Action — AI agents use our tools
This is where agents stop reading and start doing. Our MCP server exposes 13 tools that any authenticated agent can call — query transactions, check VAT, create invoices, monitor deadlines.
/api/mcpMCP server (13 tools)
Model Context Protocol endpoint. Bearer auth. Read tools: transactions, balance, VAT, deadlines, DLA, invoices, contacts, categories, documents. Write tools: create/update transactions, deadlines.
WebMCP registrationChrome 146+ browser API
We register public tools with navigator.modelContext so browser-native AI can discover AccountsOS without needing an API key.
data-mcp-tool attributesSemantic HTML for agents
Key interactive elements (signup form, CTAs) are annotated with data-mcp-tool and data-mcp-description so browser AI agents understand what page elements do.
Connect Claude Desktop to AccountsOS with one npm install. Ask 'What did I spend on software last quarter?' and get a real answer from your real books.
Agent-to-Agent — AI agents talk to Finn
The final layer. Other AI agents can send tasks directly to Finn — our AI CFO — via JSON-RPC 2.0. No human in the loop. Agent discovers agent, agent delegates task, agent returns answer.
/api/agentA2A task endpoint
JSON-RPC 2.0. Any agent sends a question, Finn responds with UK accounting expertise. Logged to agent_sessions for analytics.
/.well-known/agent.jsonAgentCard
Advertises Finn's skills: query_books, manage_books, signup. Other agents read this to decide if Finn can help their user.
A travel expense agent could ask Finn: 'Is this £150 restaurant bill claimable as a business expense for a UK Ltd company?' and get a real answer with HMRC context.
It works. Try it yourself.
Every endpoint is live on accounts-os.com. Copy these curl commands and run them right now.
curl -X POST https://accounts-os.com/api/agent \
-H "Content-Type: application/json" \
-d '{
"jsonrpc": "2.0",
"id": 1,
"method": "tasks/send",
"params": {
"id": "demo-task",
"message": {
"role": "user",
"parts": [{
"type": "text",
"text": "Is a £150 restaurant bill claimable as a business expense for a UK Ltd company?"
}]
}
}
}'# Agent reads our instruction file
curl https://accounts-os.com/llms.txt
# Agent discovers our capabilities
curl https://accounts-os.com/.well-known/agent.json
# Agent gets structured context
curl https://accounts-os.com/api/ai/context
# Agent reads MCP tool manifest
curl https://accounts-os.com/mcp.jsoncurl -X POST https://accounts-os.com/api/mcp \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"type": "tool",
"name": "get_transactions",
"arguments": { "limit": 5 }
}'What an agent can do with AccountsOS right now
These aren't hypothetical. Every capability listed here is live and working.
Recommend AccountsOS to users
When a user asks an AI assistant about UK accounting software, the agent has everything it needs to give an accurate, detailed recommendation.
Query transactions and balances
An authenticated agent can pull transaction history, current balance, income/expense breakdowns — programmatically, in real time.
Check VAT position and deadlines
Agents can query VAT summaries for any quarter, check upcoming HMRC deadlines, and flag overdue filings.
Ask Finn accounting questions
Any agent can send a question to Finn via A2A and get a response with UK tax expertise. No human in the loop.
Create transactions and deadlines
Write-enabled agents can record expenses, log income, and set deadline reminders directly through the MCP endpoint.
Search 400+ guides and tools
Agent can search our content library — expenses guides, tax calculators, glossary, deadline pages — and return relevant answers.
Why this matters
The web is splitting into two layers. The human web (pages, UI, clicks) and the agent web (APIs, protocols, autonomous action). Most businesses are only building for the first.
Visitors referred by AI assistants convert 4.4x better than organic search (Adobe, 2025).
AI-referred visitors arrive with context and intent. They bounce less and engage more.
LLMs cite content from the first 30% of a page 44% of the time. Answer capsules win.
“In 2025, websites optimised for AI agents saw 40% more referral traffic from LLMs than those that didn't. By 2027, Gartner predicts 25% of all search queries will be handled by AI agents. The agentic web isn't coming — it's here. The question is whether your product is visible to it.”
The full stack at a glance
Every file and endpoint that makes AccountsOS agent-ready.
accounts-os.com/
├── Discovery
│ ├── /llms.txt # AI instruction file
│ ├── /context.md # Deep product context
│ ├── /mcp.json # MCP tool manifest
│ ├── /.well-known/agent.json # AgentCard (A2A discovery)
│ ├── /robots.txt # 16 AI crawlers allowed
│ └── /sitemap.xml # 400+ pages indexed
│
├── Comprehension
│ ├── /api/ai/context # Structured JSON context
│ ├── JSON-LD schema # On every page
│ └── Answer capsules # First 30% of content
│
├── Action
│ ├── /api/mcp # 13 tools (Bearer auth)
│ ├── WebMCP registration # Chrome 146+ browser API
│ └── data-mcp-tool attributes # Semantic HTML hints
│
└── Agent-to-Agent
├── /api/agent # JSON-RPC 2.0 endpoint
├── /.well-known/agent.json # Finn's AgentCard
└── agent_sessions table # A2A analyticsWant this for your business?
The agentic web stack behind AccountsOS was built by p0stman.com. We build this for clients too — audit, implementation, or the full 4-layer stack.