Re:Show HN: Standard Support – a convention for AI agents to...
by nathan_f77 on Monday November 16, 2026 @06:12PM
(#50524918)
Hi HN, I'm Nathan, the founder of DocSpring [1] (an API for filling in and signing documents).
We recently put a lot of work into our MCP server and API, because more of our customers' setup work is done by their AI agents: creating templates, testing them with tool calls, and writing their integrations. We get error reports automatically, but a lot of bugs and friction never reach us unless someone tells us about it.
I asked one of my AI agents to help me write up how an agent could report a problem and get an answer back. We decided on four MCP tools (open a ticket, list, read, reply), the same operations over HTTP, and a status URL per ticket that a background script can poll. Some people set up email inboxes for their agents, but most people running Claude Code probably won't have those set up. We also decided that tickets belong to the customer, not the ephemeral agent session, so a fresh session tomorrow can pick up any replies.
A few projects already do something like this. What I couldn't find was a written-down set of decisions anyone could copy. You can consider this more of a suggestion and a reference implementation than a "standard", but I liked the name Standard Support. Nothing breaks if you implement it differently, and you're free to build your own support tools that work differently. Agents will read the tool descriptions either way.
We haven't shipped it in DocSpring yet and it's still a WIP draft. I decided to publish this early because I'd be interested in hearing your feedback. For example, what do you think about the following:
- Polling a status URL instead of webhooks. Coding agents could make this easier by letting an agent keep an MCP tool call polling in the background. There's also work on MCP Tasks and MCP Events [2].
- The default that the agent asks the user before opening a ticket. I actually lean towards inviting more reports for our own product, but I didn't want the default to flood anyone's inbox. However, we worked with a partner recently and their AI agent gave us a huge list of feedback for how we can improve our API to work better with HTML templates. There were a lot of really good suggestions and we implemented almost everything.
- The rules for when agents are on both sides of a ticket. Reports and replies are both untrusted text.
We've included a reference server and an agent skill in the repo [3] in case you would like to explore this further.
We recently put a lot of work into our MCP server and API, because more of our customers' setup work is done by their AI agents: creating templates, testing them with tool calls, and writing their integrations. We get error reports automatically, but a lot of bugs and friction never reach us unless someone tells us about it.
I asked one of my AI agents to help me write up how an agent could report a problem and get an answer back. We decided on four MCP tools (open a ticket, list, read, reply), the same operations over HTTP, and a status URL per ticket that a background script can poll. Some people set up email inboxes for their agents, but most people running Claude Code probably won't have those set up. We also decided that tickets belong to the customer, not the ephemeral agent session, so a fresh session tomorrow can pick up any replies.
A few projects already do something like this. What I couldn't find was a written-down set of decisions anyone could copy. You can consider this more of a suggestion and a reference implementation than a "standard", but I liked the name Standard Support. Nothing breaks if you implement it differently, and you're free to build your own support tools that work differently. Agents will read the tool descriptions either way.
We haven't shipped it in DocSpring yet and it's still a WIP draft. I decided to publish this early because I'd be interested in hearing your feedback. For example, what do you think about the following:
- Polling a status URL instead of webhooks. Coding agents could make this easier by letting an agent keep an MCP tool call polling in the background. There's also work on MCP Tasks and MCP Events [2].
- The default that the agent asks the user before opening a ticket. I actually lean towards inviting more reports for our own product, but I didn't want the default to flood anyone's inbox. However, we worked with a partner recently and their AI agent gave us a huge list of feedback for how we can improve our API to work better with HTML templates. There were a lot of really good suggestions and we implemented almost everything.
- The rules for when agents are on both sides of a ticket. Reports and replies are both untrusted text.
We've included a reference server and an agent skill in the repo [3] in case you would like to explore this further.
[1] https://docspring.com
[2] https://modelcontextprotocol.io/community/working-groups/tas...
[3] https://github.com/DocSpring/standardsupport