I build the tool layer that agents call, and I've had one in production long enough to know where it breaks.
At Chronicle I built our hosted MCP server end to end — eight tools that Claude and ChatGPT authenticate into over WorkOS OAuth with JWT and async polling, running on ECS. Also our first public REST API and the developer docs, a workspace-scoped API key system for tenant isolation, and an npm node that had to pass n8n's strict-mode security review before listing. On the interface side, an AI-generation UX with streaming tool-call visualisation, cancel/abort mid-flight, and live document editing over CRDTs.
What I've learned is that agent failures are boring and specific rather than clever. I lost four days to what looked like an auth error and turned out to be a tool declared without an input schema, silently changing the signature the SDK used to call my handler. Four theories before I read the error string properly. I write the theory down before chasing it now.
I also shipped Recall (<a href="https://recallapp.online" rel="nofollow">https://recallapp.online) solo in about two weeks — notes into flashcards over MCP, with FSRS scheduling. Design, backend, payments and launch were mine. #22 on Product Hunt. Before Chronicle I owned a Shopify-to-custom-platform migration end to end, including ingest pipelines from three sources into Postgres.
~4 years. Looking for a remote backend, full-stack, or agent-infrastructure role where I own things end to end rather than picking up tickets.
Open to full-time or contract.
tushar-nath · · focus · HN ↗
Remote: Yes — remote only. Full overlap with European hours, plus US-East mornings.
Willing to relocate: No
Technologies: TypeScript, Node.js, React, Next.js, PostgreSQL (Drizzle, Neon), MCP, OAuth 2.0/2.1, JWT, AWS (ECS, EC2, Lambda, SQS, S3), Docker, CI/CD, Redis, XState, Yjs/Liveblocks CRDTs, Vitest, Stripe/Polar, Claude Code, Cursor
Résumé/CV: <a href="https://tusharnath.online" rel="nofollow">https://tusharnath.online
GitHub: <a href="https://github.com/tushar-nath" rel="nofollow">https://github.com/tushar-nath
LinkedIn: <a href="https://linkedin.com/in/tusharnath10" rel="nofollow">https://linkedin.com/in/tusharnath10
Email: tusharnath10@gmail.com
I build the tool layer that agents call, and I've had one in production long enough to know where it breaks.
At Chronicle I built our hosted MCP server end to end — eight tools that Claude and ChatGPT authenticate into over WorkOS OAuth with JWT and async polling, running on ECS. Also our first public REST API and the developer docs, a workspace-scoped API key system for tenant isolation, and an npm node that had to pass n8n's strict-mode security review before listing. On the interface side, an AI-generation UX with streaming tool-call visualisation, cancel/abort mid-flight, and live document editing over CRDTs.
What I've learned is that agent failures are boring and specific rather than clever. I lost four days to what looked like an auth error and turned out to be a tool declared without an input schema, silently changing the signature the SDK used to call my handler. Four theories before I read the error string properly. I write the theory down before chasing it now.
I also shipped Recall (<a href="https://recallapp.online" rel="nofollow">https://recallapp.online) solo in about two weeks — notes into flashcards over MCP, with FSRS scheduling. Design, backend, payments and launch were mine. #22 on Product Hunt. Before Chronicle I owned a Shopify-to-custom-platform migration end to end, including ingest pipelines from three sources into Postgres.
~4 years. Looking for a remote backend, full-stack, or agent-infrastructure role where I own things end to end rather than picking up tickets. Open to full-time or contract.