guest@gurusup > /changelog

Changelog: Agents

What we've shipped in Gurusup Brain, newest first.

1. AgentsOctober 8, 2026

   Your agent asks Gurusup Brain with a subject and gets full evidence

   The ask_company_memory tool took more than ten parameters for every question. Its answers also came with a confidence label and a list of gaps that trimmed the evidence, so the agent ended up second-guessing Brain instead of reading what it found.

   Now the tool takes a subject (the customer, project or person the question is about), a question and an optional activity_id. It returns the evidence in full, passages and claims included, and leaves the judgement to your agent, which knows what it is trying to answer.

   • └Answers no longer include a confidence label or gaps, and new queries in Queries no longer show them.
   • └Answers can be longer, because nothing is cut.
   • └Calls without subject are rejected. Refresh the tool schema in your client and pass subject, or null when the question is not about anything specific.

   With the subject named up front, a question like "what is the status of the renewal?" is tied to the right customer, and your agent gets all the evidence to give a complete answer.

   MCP reference →
2. AgentsOctober 4, 2026

   Connect your agent with one card per tool

   Gurusup Brain is meant to be used from the AI tool you already work in. Connecting it used to be one generic step with a server address, and each tool wants that address in a different place.

   Now Connect your agent has a card for each client, with the exact steps for that tool: Claude Desktop, Claude Code, Codex, ChatGPT, Cursor, VS Code and Other MCP client.

   • └Cursor and VS Code install in one click.
   • └Claude Code and Codex give you one command to paste in your terminal.
   • └ChatGPT adds it as a custom connector in Settings, and Claude Desktop as an MCP server in the desktop app.
   • └Other MCP client gives you the server URL to paste into any client that supports remote MCP.

   Whoever sets it up only needs the card for their tool, and once it is connected they can ask about your company's knowledge without leaving the conversation they are already having.

   MCP reference →
3. AgentsSeptember 28, 2026

   Two new tools let your agent tell us when Brain falls short

   The agent using Brain had no way to tell us what went wrong. Two new MCP tools fix that:

   • └report_issue: a failure the agent can point to. A tool error, information the user says exists but Brain didn't find, or an answer that contradicts its own evidence.
   • └send_feedback: an improvement when nothing failed, sent once at the end of the task.

   Reports are linked only to the caller's own activity, by id, so we can trace what happened without reading anything else in your organization.

   MCP reference →