HivemindOS manual

Hive Skills

Hive skills are HivemindOS’s own workflows. They teach agents how to use the hive’s memory, tools, apps, machines, and projects. Setup installs them into your shared brain automatically, at Skills/<skill-name>/SKILL.md.

Catalog

Skill	Purpose
`photo-library-keyworder`	Indexes large image folders with the desktop Photo Library, sends unclear cases to human review, and exports Lightroom-compatible XMP sidecars.
`harness-engineering`	Runs fixed-worker comparisons of context, tools, and instructions, then records a keep, revise, or remove decision. Adapted from Ryan Lopopolo’s CC BY 4.0 harness-engineering project.
`engineering-discipline`	Picks an engineering path that fits the risk: real baselines, tests, focused verification, independent review, and rollback.
`bankr-skill-deployment`	Chooses the correct Bankr skill, app, x402 endpoint, or always-on host, then packages, publishes, installs, and verifies it.
`bankr-app-development`	Creates, tests, publishes, shares, and repairs Bankr apps, including permissions, schedules, secrets, x402 actions, live checks, and rollback.
`create-zero-human-company`	Creates a company from a business goal or repository with a crew, goal, products, budgets, and approval rules. Autonomy stays off until you launch it.
`operate-zero-human-company`	Runs an existing company in priority order: setup blockers, expiring spend approvals, decisions that need you, budget review, and launch or pause. It never skips approval gates.
`hive-assimilate`	Searches your brain, projects, private indexes, and public GitHub before any new software is built.
`hive-pulse`	Last-30-days signal briefs from Reddit, X, YouTube, TikTok, Hacker News, Polymarket, GitHub, and the web, using a pinned MIT-licensed `last30days` engine.
`reddit-thread-reader`	Reads any Reddit link you share, then shows, summarizes, or judges the sentiment of the post or thread, on desktop and phone.
`hive-quant-research`	Research-only strategy testing with simulation, independent statistics, overfitting checks, and approved scheduled runs.
`hive-capability-search`	Finds the tools, skills, optional playbooks, apps, agents, credential names, and delivery channels that fit a task.
`hive-remote-capability-use`	Uses apps and tools on your other machines that capability search found, with private file transfer, output checks, and approval for side effects.
`remote-machines`	Runs, schedules, moves, and removes automations on your always-on machines, and sets their wallet access. Every change needs Touch ID, a passkey, or your password.
`hive-skill-fusion`	Turns a capability request into a reusable shared-brain skill.
`hive-workflow-fusion`	Combines skills, apps, agents, and tools into a multi-step workflow.
`hive-aeon-fusion`	Turns a reusable workflow into a scheduled AEON duty when that fits.
`hive-brain-memory`	Recalls typed memory, keeps routine receipts separate, reviews mined patterns, and updates stale memories while keeping their history.
`hive-brain-compiled-wiki`	Writes durable findings as compiled wiki pages (a page an agent writes waits for your review), lets agents search/query the compiled graph through MCP, and repairs wiki health.
`hive-skill-autoresearch`	Improves an installed skill by testing four variants against a baseline and returning a winning change for you to approve.
`wrapup`	When you ask to wrap up a session, saves reviewed durable facts to Shared Brain Memory and adds a short redacted summary to your NotebookLM AI Brain notebook.

Supporting Hive Search Commands

Memory search uses the hive-brain command. There is no separate skill called hive-find or hive-search.

hive-brain answer "query"
hive-brain recall "query" --scope full-vault --limit 8
hive-brain evolve --memory-id mem-... --content "Updated durable memory"

hive-brain answer returns a short grounded answer.
hive-brain recall returns a list of hits.
hive-brain evolve replaces a stale memory and keeps the old version as history.

Both search commands use the running app first and fall back to local vault search when it is closed. For the full memory policy, see the hive-brain-memory skill. For compiled wiki knowledge, agents use the hive-brain-compiled-wiki skill, which searches compiled pages first.

To find the right tool for a task, setup also installs hive-capability-search:

hive-capability-search "task or goal"
hive-capability-search --json --kinds skill,tool-schema "task or goal"
hive-capability-search --live "task or goal"

It searches a local index and works without the dashboard. --live asks the running dashboard first for fresher app details.

When the app is running, a result can also show a record line: how often that tool worked when it was used on this computer in the last 30 days, how long it usually took, and when it last worked. A reports line counts the agents that reported a problem with it. Agents use these to choose between tools that fit, and the same skill tells them how to retry after a failure, when to say what something costs, and how to report a problem with a tool once and carry on.

Agents without shell access, such as phone-hosted agents, don’t need the command. They use the capability context HivemindOS already gave them or an MCP tool instead.

Hive Assimilate

hive-assimilate runs before an agent builds software, so it reuses what exists. It searches in this order:

Sources you pinned, such as a repo, pull request, note, or skill.
Your shared brain.
The current workspace’s files, docs, and tests.
Your other projects, when relevant.
Local and private indexes.
Public GitHub, sorted by stars.

It writes a log in the project (ASSIMILATION_LOG.md) so you can see what was searched, reused, or rejected.

Hive Pulse

hive-pulse gives you a brief of what people are saying publicly about a topic over the last 30 days. It works right after setup and is read-only.

Reddit, Hacker News, Polymarket, GitHub, and the web work with no extra setup. Adding your own keys expands coverage to X, YouTube, TikTok, Instagram, Threads, and Pinterest. Agents check keys by name only and never print their values.

Reddit Thread Reader

Paste a Reddit link (including app share links) and say what you want:

“Pull this thread” shows the post and comments.
“Pull this post” shows just the post.
“Summarize this thread” or “What’s the sentiment?” gives an answer based on the comments.

In the phone app, the thread is attached to your message, so it works with any model. Very large threads keep their top-scored comments. Posts in private communities can’t be read.

Bankr skills

bankr-skill-deployment decides how a capability should run with Bankr: as a per-wallet skill, a paid x402 endpoint, a Bankr app, or a separately hosted always-on backend. It then publishes and verifies it without exposing secrets.

bankr-app-development builds and releases Bankr apps. Each release is checked live and can be rolled back, and repeating a release changes nothing.

Optional and third-party skills
Shared Skills
Hive Fusion
Token And Cost Savings: how hive-brain, hive-capability-search, hive-assimilate, and karpathy-guidelines cut model costs.