Why I'm Building agentplatform
Every AI agent built for a business needs the same pieces. A way to receive messages. An LLM to process them. A knowledge base to ground the responses. Maybe a CRM integration. Maybe guardrails to filter the output.
And every developer building these agents wires those pieces together from scratch. Every time.
That's the problem I'm trying to solve.
The build-from-scratch trap
If you're a developer building a chatbot for a client today, here's what your week looks like:
- Wire up an LLM provider. Manage API keys, handle streaming, build retry logic.
- Build the chat interface. WebSocket or SSE, session management, message history.
- Connect the knowledge base. Chunk documents, build embeddings, set up vector search.
- Integrate the CRM. OAuth flow, API wrappers, data mapping.
- Add guardrails. Content filtering, PII detection, response validation.
- Host all of it. Containers, load balancing, monitoring.
You wanted to build an agent. Instead you built infrastructure.
Next client, next agent — you do it all again. Maybe you copy-paste from last time. Maybe you build an internal library. But it's still custom code, still your hosting, still your maintenance burden.
What if agents were assembled, not built?
The pieces are the same every time. What changes is the combination: this client needs web chat + Claude + HubSpot. That client needs Telegram + GPT-4 + Shopify. The pieces are interchangeable. The wiring is the only variable.
So what if the wiring was a config file?
agent "acme-support" {
channel "web_chat" {
source = "agentplatform/channel-webchat"
}
llm "anthropic" {
source = "agentplatform/llm-anthropic"
model = "claude-sonnet-4-20250514"
}
plugin "knowledge-base" {
source = "agentplatform/rag"
hook "pre_call" {}
config {
data_source = "https://docs.acme.com/sitemap.xml"
}
}
plugin "hubspot" {
source = "agentplatform/crm-hubspot"
hook "tool_call" {}
}
}
That's a complete support agent. Web chat, Claude, knowledge base, HubSpot. One file. Push it, the agent is live.
Why HCL
If you've used Terraform, you know HCL. It's declarative, composable, and designed for infrastructure-as-code. It's not YAML — which matters more than you'd think when you're nesting configurations three levels deep.
HCL gives you things a visual builder can't: version control, diffs, code review, composability. You can template an agent definition and customize it per client. You can check it into git and track changes over time.
But you don't have to write HCL if you don't want to. The dashboard has a form-based UI that reads and writes HCL behind the scenes. Same artifact, two interfaces.
Plugins, not integrations
Most agent platforms build their own integrations. Botpress has 190 of them. But they're all built by Botpress. If you need one that doesn't exist, you wait for them to build it — or you leave.
agentplatform takes the opposite approach. Plugins are gRPC services that anyone can build, in any language. Build a HubSpot integration in Python. Build a guardrail in Rust. Publish it to the marketplace. Set your price per call. Other developers add it to their agents with three lines of config.
This is the difference between a product and a platform. A product has features. A platform has an ecosystem.
What makes plugins different from skills
The AI ecosystem already has skills (markdown prompt files) and MCP servers (local tool providers). Plugins are something else. (For a deeper comparison, read Skills vs Plugins vs MCP Servers.)
Skills change how an agent thinks. They're prompt instructions. No logic, no API calls, no state.
MCP servers add tools an agent can call. Powerful, but you manage the hosting and lifecycle yourself.
Plugins hook into the entire LLM call lifecycle. They can intercept requests before they reach the LLM, transform responses before they reach the user, trigger webhooks, enforce policies, connect to external systems with server-side credentials the agent never sees. And they're hosted — you publish, we run it.
Where we are
The marketing site is live. The engine isn't built yet. I'm starting with a small group of early access developers to get the plugin API and HCL format right before opening up.
If you build AI agents for businesses — or want to — I'd like to hear from you. What plugins would you need first? What integrations are you tired of rebuilding?