HivemindOS manual
Agent-Native Workflows
These features let agents coordinate work through HivemindOS while keeping everything they do visible and reviewable. Records live in your shared vault when it is available, or in local HivemindOS state otherwise. They improve visibility; they never grant an agent extra authority.
MCP tools for agents
The HivemindOS MCP server gives any MCP-capable agent the same coordination tools the dashboard uses:
| Tool | What it does |
|---|---|
work_board |
List, create, claim, heartbeat, comment on, complete, fail, block, or promote Work board tasks. |
queen_bee |
Check Queen Bee, queue work, break a PRD into tasks, or run Queen Bee flow templates. |
work_event |
Define event triggers and publish events that turn into ordinary Work board tasks. |
request_human_approval |
Create a Needs You card for a human decision. It does not approve anything itself. |
spawn_subagent |
Start self-contained work in a fresh child session and return a job id right away. |
check_subagent |
Check a child’s status without waiting. |
await_subagent |
Wait for a child’s result when the parent needs it. |
run_subagent |
Start a child and wait for it in one call. Supports standard MCP Tasks. |
Work still lands on the Work board, and Queen Bee still handles routing and receipts.
Wallet and trading tools (send_usdc, dex_swap, stock_trade) are marked as critical and need their exact confirmation token. The dashboard still runs its own spending, wallet, and settlement checks.
Subagents and context isolation
Use spawn_subagent when a child needs substantial tool work and the parent can keep going. Check on it with check_subagent, and call await_subagent only when you need the result. run_subagent does both in one step; MCP clients on protocol 2025-11-25 get it as a background task that notifies on completion.
How isolation works:
- The child doesn’t see the parent’s conversation, so the task must stand on its own.
- The parent gets only the child’s final response, within a 1,000–12,000 character budget you can set. Intermediate steps stay with the child.
- The requesting agent isn’t picked as the child when another eligible worker exists.
- Results are hypotheses until the parent checks their evidence.
This keeps the parent’s context small, but it doesn’t prove lower total usage: the child spends its own tokens.
Children can run on HivemindOS, Hermes, Codex, Claude Code, OpenCode, OpenHands, Aider, Claw, or OpenClaw. The chosen runtime must be ready on its machine; HivemindOS won’t swap in a different one. A child job fails clearly if its machine restarts mid-run. Waits are capped at 20 minutes; longer work belongs on the Work board.
Cross-session publishing
An explicit request to commit, merge, land, publish, or push work across all sessions starts immediately from the repository’s current Git state. HivemindOS captures the shared checkout, attached worktrees, local branches, and unpublished commits without contacting, inspecting, or waiting on other conversations.
It keeps captured changes as normal commits and branches, runs proportionate checks, scans for secrets, merges safe local branches, and pushes without force. Other repositories are not touched.
After the first push, HivemindOS makes one immediate second Git pass to catch changes that arrived during publishing. Anything it can’t include without a destructive conflict resolution is reported in the conversation.
Dashboard pins
Pins let you leave feedback on any part of the dashboard.
- Open the pin overlay and click the part of the screen you mean.
- Write a note and save it.
- Resolve, archive, or delete the pin later, or send it to the Work board as a task.
Pins are saved by HivemindOS, not in browser storage. A pin never changes anything by itself.
Shared Brain review queue
Agents propose changes to durable shared knowledge instead of making them silently. The review queue holds new memories, memory updates, skill edits, instruction edits, and jobs.
Open the Memory view to approve, reject, or apply proposals. Approved memory proposals are written through the normal Shared Brain Memory path. Skill, instruction, and job proposals are applied by hand for now.
Context X-Ray
Context X-Ray records what an agent run could see: the sources, why each matched, a short snippet, and an estimated token count, with sensitive values redacted. The Memory view uses it to show why an agent saw a particular skill, tool, app, runtime, doc, or file. X-Ray records describe context; they are not trusted source data.
Visual plans and recaps
Visual plans and recaps give a quick map of a piece of work: a summary, file tree, diagrams, wireframe notes, a diff summary, and risk notes.
- Queen Bee creates a visual plan when it queues a Work board task or breaks down a PRD, linked back to that task.
- Local code work can produce a recap of changed files, risky areas, and what to review.
Plans and recaps describe work. They never approve a deployment, spending, credential access, or destructive changes.