HeroMind for agents · MCP

The context layer
your agents share.

One MCP server with your tasks, your library and your record behind it. Connect ChatGPT, Claude, or the agents you run yourself, and they read before they act, file work with a reason attached, and propose instead of overwrite.

ChatGPTClaudeClaude Code HeroAgentsHermesGrokany MCP client
The contract

Three rules every connected agent works under.

1 READ FIRST

Start from what I've decided

Before a run, an agent asks for a context pack sized to a budget: who I am, standing preferences, the decisions relevant to the job, what happened recently, today's plate. My words, verbatim, with provenance.

2 ACT INSIDE IT

File work, with the why

A task carries a description of why it exists and a link back to the run that made it. A dedupe key means a re-run never files it twice. Findings that don't need action go on a daily card, not the task list.

3 LEAVE A RECORD

Propose, and log the run

A decision an agent thinks I made arrives as a proposal on the Desk. An episode says what ran and what changed. Nothing becomes true about me without my yes.

This is the planner's rule extended to other people's software. It advises. You decide. Each agent gets its own revocable token, and anything that would change my record goes to my review queue, the Desk, first.

The tools

What an agent can read, what it can write, what it can't touch.

ToolWhat it does
get_contextA pack for a task: { query, budget }. Ranked by meaning and keywords over my decisions, preferences and beliefs, plus recent episodes and today's tasks. Deterministic and under a second, so "why did the agent think that" is always answerable.
library_searchHybrid search over articles and sources, full-text and vector fused. The same search the Library's own search box uses.
library_read_article
library_read_source
Read a wiki page, an editorial or a source in full, with its footnotes.
list_tasksOpen and completed tasks, with updated_since so a routine sees what changed since it last looked.
create_personal_task
create_work_task
File a task into the right space, with description, source_url and an external_ref that makes the call idempotent.
library_add_sourceFile a page, a thread, a transcript or a note into a topic, or into the Inbox for the librarian to place. Origin is recorded.
upsert_daily_noteThe morning card on my planner. Idempotent, so a re-run replaces rather than duplicates.
capture_context
add_decision
Propose a decision, preference or belief in my words. It waits on the Desk, and accepting it can retire the one it replaces.
add_episode
get_timeline
Log what a run did; read what happened on a day. The timeline feeds the weekly reflection.

Always

  • Every call is authenticated with a token I issued, per client, revocable.
  • Every call is logged: which client, which tool, how long.
  • What an agent files says who filed it, everywhere it appears.

Never

  • Change a current decision or belief directly.
  • Read a source I've marked private.
  • Delete anything.
  • Write to my own notes page. The morning card is theirs; the Notes tab is mine.
A morning, as it actually runs

Overnight work arrives with a reason attached.

02:00 · Hermes→LibraryFour briefs and the X threads that matter. It searches first so nothing is filed twice. 33 sources in a typical fortnight.
04:00 · the librarian→Wiki · DeskCompiles the topics that moved, writes an editorial where one is due, proposes people and decisions.
06:00 · Chief of Staff→PlannerReads the context pack and last night's completions, then files one morning card: what changed, what needs me, the top three.
Email triage→TasksReal actions with a due date and the email linked. Drafts wait for me; sending is mine.
Me, with coffee→The DeskA yes or a no on what the night proposed. Silence for two weeks is a no.

I run HeroAgents and Hermes. The flow is the same with any runtime that speaks MCP; only the names change.

Connect

One endpoint. OAuth. No secrets to paste.

The server is https://mcp.heromind.ai/mcp. It speaks OAuth 2.1 with PKCE and registers clients dynamically, so the assistants that support custom connectors set themselves up.

ChatGPT

Settings → Connectors → Create. Paste the endpoint, approve the consent screen in HeroMind. Add one line to your instructions: start with get_context.

Claude

Settings → Connectors → Add custom connector, same endpoint. Works in Claude.ai and Claude Desktop.

Claude Code and other agents

Add the endpoint as an HTTP MCP server and authorise once. For runtimes without a browser, issue a token in HeroMind and pass it as a bearer.

// what the Chief of Staff asks for, every morning
get_context({
  query:  "morning briefing: what needs me today, open loops, how he wants the day framed",
  budget: 6000
})
// → header · standing preferences · relevant decisions · recently · today's plate

The full API reference ships with the beta invite. Tokens are per client and revocable from Settings; the usage log shows every call.

Where this sits

Agent memory is a crowded shelf. This is the other half of it.

Mem0, Zep, Hindsight and Garry Tan's gbrain are serious memory stores for agents, and on the graph and on scale they go deeper than HeroMind does. Their design centre is the agent. Mine is the person the agents work for.

So the differences are about where the human sits: I read first and the agents read second. A belief about me carries a status and can be superseded. The nightly work runs as code rather than as a prompt an agent may or may not follow. And the whole thing sits next to my tasks and my day, so a source can become an editorial, an editorial a task, a task a line on the timeline.

If you want a shared brain for a fleet of agents and nobody will read it, use one of theirs. If you want to read it, this is built for you.

Bring the agents you have

Give your agents one place to read from and file into.

Free during beta · MCP access included · tokens per client, revocable anytime