How to use Claude to build a local MCP server
CasperAI’s creator used Claude Code to generate about 80% of its TypeScript, including the MCP server and SQLite schema. The author retained architecture, debugging, documentation and edge-case testing.
news.ycombinator.com
Prompt to try
Editorial adaptation, not the source author’s exact prompt. Replace bracketed inputs and review tool permissions first.
Help build [local engineering-context MCP server] using TypeScript, Node.js, SQLite and the MCP SDK. Implement connections to [development tools], PII redaction, a SQLite schema and regex-based code-reference extraction with filesystem resolution. Keep data local. Identify false positives and definition-versus-call-site limitations for my architecture review and edge-case testing.
Reach people exploring AI agents
See placements →What happened
The author built CasperAI, a local SQLite-backed MCP server connecting engineering context to source code, with roughly 80% of its TypeScript generated by Claude Code. Regex mapping retained acknowledged accuracy limitations.
Reported by chose166 on Hacker NewsHow the agent did it
- Used Claude Code to write the MCP server.
- Used Claude Code to implement PII redaction and regex patterns.
- Used Claude Code to create the SQLite schema.
Where a person stepped in
The author handled architecture, debugging and documentation, understood the trade-offs and tested edge cases.
A person guided the agent closely as it worked.
The creator of CasperAI used Claude Code to help build a local MCP server that connects engineering context across Slack, GitHub, Jira and Notion with source code. The author reported that roughly 80% of the TypeScript was generated with Claude Code. They specifically attributed the MCP server, PII redaction, regex patterns and SQLite schema to the agent.
The author retained architecture, debugging and documentation work, and said they still had to understand the design, make trade-offs and test edge cases. The stack comprised TypeScript, Node.js, SQLite and the MCP SDK. All data stayed local in SQLite.
For code mapping, the implementation extracted code references from natural language using regex-based pattern matching, then resolved them through filesystem traversal. Search brought together a Slack thread, file location and PR discussion. The author chose regex over AST parsing to work across more than 15 languages without language-specific parsers or parsing entire files.
They also acknowledged possible false positives and an inability to distinguish definitions from call sites. Tree-sitter was a future plan, not a completed component. Who it's for: Developers building local tools to connect engineering discussions with code who can own architectural decisions and test generated implementations.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Hacker News · chose166
Frequently asked questions
Which parts did Claude Code write?
The author attributed the MCP server, PII redaction, regex patterns and SQLite schema to Claude Code, and estimated that about 80% of the TypeScript was generated with it.
What work did the author retain?
The author handled architecture, debugging and documentation, made implementation trade-offs and tested edge cases.
What were the code-mapping limitations?
The regex approach could produce false positives and could not distinguish definitions from call sites. The author planned future tree-sitter integration for more precise AST-based resolution.
More details
Remaining steps
- Handled architecture, debugging and documentation personally.
- Tested edge cases and evaluated the regex code-mapping trade-offs.
About this account
Claude Code generated approximately 80% of the TypeScript, including the MCP server, PII redaction, regex patterns and SQLite schema.
Platform: Hacker News. Attribution: https://news.ycombinator.com/item?id=47048893. Prompt: Suggested (written by Agents In Use, not the author's). Sourced by the Agents In Use pipeline.
Credit: chose166 on Hacker News
Source published 17 Feb 2026
Source checked 14 Oct 2026
Tools and topics
Claude Code · TypeScript · Node.js · SQLite · MCP SDK · Slack · GitHub · Jira · Notion
MCP · Local storage · Code mapping
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.