Librarian OS

Why this OS

What it is for

A founder, operator, or another Agentik OS that needs the compressed intelligence of a field, with evidence, structure, and the right to go deeper.

Asking a model to summarize a book or a topic produces a confident essay. It does not record what was actually read, keep disagreements, or feed another Operating System.

The unit of work is a Knowledge Pack with a coverage ledger. Uninspected chapters stay unnamed. Builder can call this OS for domain intelligence instead of inventing an architecture from pretrained memory alone.

Tools you can connect

Curated for this OS. Not a Composio catalog dump. Nothing in this list is a live vendor on mcp.agentik-os.com except Agentik-OS itself.

• Drive

  Connectable / planned

  Google Drive or a similar store for user-provided books and docs. Not pre-wired.

• Docs

  Connectable / planned

  Notion or Docs as a place to reopen a Knowledge Pack. Connectable, not installed here.

• Repo

  Connectable / planned

  GitHub when Builder asks for OS intelligence. A pack is not a compile.

Orchestration

Director

Librarian Director: hold the Research Contract, dispatch the smallest useful NanoTeam set, keep provenance, stop when marginal gain is low.

NanoTeams

Seven NanoTeams: Research, Reading, Intelligence, Knowledge, Learning, OS Intelligence, Governance. Defined is not concurrent. Deep Book, Bestseller Scout, and Teaching Style are compiled specialists.

Workflows

Four compiled workflows. Deep research, book intelligence, learn path, OS intelligence. Each ends on an owner gate.

Gates

owner-gate waits for you on every compiled path. A deep mission is not done until QA passes or the gaps are explicit. RELEASE is not this OS.

Objects

Research contract, source, book intelligence, claim, knowledge pack, coverage ledger, learning profile, OS intelligence pack.

Methods

• Hold the contract, dispatch, refuse unread claims
• Write a purpose-driven Research Contract
• Find candidates without ranking the canon
• List bestsellers as a popularity signal, not a canon
• Rank a representative canon, including opposing schools
• Record full, partial, or unavailable access
• Map structure only where access permits
• Analyze a chapter only if it was inspected
• Walk one inspected book chapter by chapter
• Extract concepts with provenance
• Write claims as ledger objects
• Check that sources support claims
• Distill without fake consensus
• Keep contradictions visible
• Turn knowledge into procedures with conditions
• Write typed relations that were observed
• Keep global, user, OS, and mission scopes apart
• Publish the coverage ledger and passport
• Adapt the pack to the Learning Profile
• Store how this person learns
• Ask once how to explain. Keep it until they change it
• Test recall. Reading is not learning
• Return a Domain Intelligence Pack for Builder
• Recommend skills. Do not implement them
• Audit hallucination and provenance before the gate

You get

• A Research Contract before a synthesis
• Access state on every source: full, partial, or unavailable
• A Knowledge Pack with disagreement still attached
• A Domain Intelligence Pack Builder can consume

It will not

• It will not invent chapter 7
• It will not average disagreement into fake consensus
• It will not reconstruct a copyrighted book
• It will not compile or release another OS

Honesty

• This is a preview OS. The package compiles. The compiler refuses RELEASE. Your host runs the model when it calls MCP.
• Twenty-five agents are defined on the organization graph. A mission activates a subset. Defined is not mobilized. Mobilized is not concurrent.
• A Knowledge Pack is not proof that every important book was read. The coverage ledger is the honest part.
• A reviewed pack is not a certified literature review. We do not claim 248/248 PASS.

How a run looks

The user path for this OS. As many steps as the run needs.

Workflows

Nodes and edges from the compiled package. Nothing decorative.

Planner writes the contract. Teaching Style asks once how to explain. Scout, Bestseller Scout, and canon find sources. Access records what was actually available. Reading and Deep Book extract only inspected chapters. Debate keeps disagreement. Synthesis, graph, coverage, pedagogy, and practitioner write the pack. QA records the shape. You approve.

Result synthesizer · Completes at owner-gate

Agents

25 agents.

Before / after

A day, a week, a month, a year. Honest deltas. No invented multiple.

How to use it

Unlock ($19.99 once, lifetime, or a plan ($19 / $99 / $199) that includes official OS.) then install on the host you already pay. Talk to the Director. Call a team. Approve named artifacts. The host pays tokens.

list_agents({ osId: "librarian-os" })

ask_director({ osId: "librarian-os", prompt: "Open a deep research mission on [topic]. Write a Research Contract. Verify source access. Do not pretend a book was read." })

ask_team({ osId: "librarian-os", teamId: "source-access", prompt: "For this source, record full, partial, or unavailable. Name the sections actually inspected." })

approve({ jobId: "<job-id>", approved: true })

1. 01

   Unlock before you install

   Buy Librarian OS for $19.99 once, lifetime, or use a plan that includes official OS ($19 for three official OS, $99 for ten, $199 unlimited). Payment is confirmed before access opens. A success page never grants access.

2. 02

   Install on the host you already pay

   MCP is the HTTP connector (Cursor, Claude Code, Codex, Hermes). CLI is the repo commands. Installing does not start a research mission and does not spend anything. Tokens stay on that host.

3. 03

   Talk to the Director

   Name the topic, the book, or the OS you want researched. Ask the Director to open a Research Contract. The Director supervises, dispatches NanoTeams, and keeps provenance. It does not invent a chapter.

4. 04

   See the roster

   Call list_agents for this OS. You should see the Director plus twenty-four specialists across research, reading, intelligence, knowledge, learning, OS intelligence, and QA. A run mobilizes a path, not a headcount contest.

5. 05

   Call a team when you need one lens

   Use ask_team with an agent id (source-access, debate, pedagogy, os-intelligence) or a team id (team-research, team-reading, team-intelligence, team-knowledge, team-learning, team-os-intel, team-qa). Human labels do not resolve. Operational work still goes through jobs, not a mention loop.

6. 06

   Approve the Knowledge Pack

   The compiled graph records the pack. You approve, or you send it back. approve binds an actor, an action, a digest, a scope, and an expiration. A material change invalidates the approval.

7. 07

   What the host pays

   You unlock the organization, not tokens. Claude, ChatGPT, Codex, or Hermes bills the host. Agentik does not run a hidden model behind your back.

Install

MCP is the HTTP connector. CLI is the repo commands.

{
  "mcpServers": {
    "Agentik-OS": {
      "url": "https://mcp.agentik-os.com/api/mcp"
    }
  }
}

Other OS

Creation

Content OS

An editorial system for ideas, briefs, assets, and approvals, then an authorized package, with a human gate before anything leaves the building.

AAgentik

9

Content OS

Operations

Growth OS

A growth department: brief, audience, offer, campaign, experiment, and independent critique, without publishing or spending for you.

AAgentik

48

Growth OS