This MCP server gets smarter as your team uses it, via session_id task memory
h4xz13 · reddit · 2026-09-04
Developer h4xz13 describes how most MCP servers are stateless: every time an agent fumbles parameters or circles through tool calls, that experience is lost and users repeat the same failure loop.
Approach: treat tool calls as part of a task—ask the agent to pass the same sessionid for related calls so the MCP server understands intent and execution path. This gets much easier with code-mode MCPs, where the initial search tool reveals intent. Over time, sessions reveal which tools specific tasks need, and the server can return the right tool plus corrective instructions—effectively letting the MCP steer agent behavior. It works across any harness/runtime using your MCP. The author also built subagents that route search queries to specific apps per user.
The product is Sequel (sequel.sh), unifying GA, Search Console, Amplitude, Postgres, MySQL and other marketing/revenue/finance data into a growing 'company brain.'
More from coding & agent
- youtube-skills: open-source toolkit giving AI agents YouTube transcripts and search — tom_doerr · 2026-09-04
- Anthropic explores Function Hooks, a new way to extend Claude Code, seeking feedback — EricBuess · 2026-09-04
- Go team ships goroutine leak profiler that detects blocked channels in production — rseroter · 2026-09-04
- Catch AI launches a $99/mo executive admin agent that proactively saves $400 on hotels — eyishazyer · 2026-09-04
- Agent Token Breakdown: Only 50 of 1,300 Tokens Are Orchestration, Cutting Incident Cost to $0.001 — kalyan_kpl · 2026-09-04
- A Bayesian lens on agents: prior, action, evidence, posterior decision loops — Senior_Disaster_7307 · 2026-09-04