MCPmemory · skills · handoffs Levirge Brain

One brain for all your agents & users.

Give Claude, Codex and any MCP client — and the people working beside them — shared knowledge, reusable skills, and a reliable way to hand work to each other.

Shared across every MCP client Hosted or cloud models Update one skill for every agent
levirge brain your workspace · MCP
Levirge Brain — the entity browser with demo data: an Ask bar, Skills / Entities / Docs / Metrics tabs, a tracked-entity sidebar, and a cited per-entity summary
19
MCP tools from one server — knowledge, skills, handoffs
~6 s
per document to categorise — measured Aug 2026, local LLM
0
third-party model calls in hosted mode
The problem

Knowledge & Learnings become Siloed

Every session captures knowledge and learnings — and today they stay trapped where they happened: one session, one vendor, one machine, one person. Brain shares them across all of it — over MCP.

Animated diagram: knowledge captured in one session grows into a shared graph — across sessions, agent vendors, machines and team members — with Levirge Brain connecting all of it. Click to view full screen
What's different

A work-continuity system, not a memory store

Most memory products store what was said and hand it back. Brain also holds what your agents are doing — the session they're in, the todos they owe, the skills they follow, the handoffs they left each other — beside the knowledge itself. Four things carry that.

01

Memory and work in one substrate

Shipped

Knowledge is only half of continuity. work_session, todo, skill, handoff and reminder are tools on the same server as the knowledge base — so an agent picking up cold can ask what it was doing, what's still open, and what someone left it.

  • What was I doing — work sessions
  • What's still owed — todos
  • How we do this — versioned skills
  • What another agent left me — handoffs
  • What comes due later — dated reminders
02

Scoped per project, not per person

Shipped

Knowledge is partitioned by Vault — one per project, product or client. Memory tied to a user account smears every project you touch into one pile; a Vault keeps one client's work out of another client's answers, and lets a whole team share the one that matters.

  • A Vault per project, not one bucket per user
  • Search, entities and skills scoped to it
  • Personal Vaults stay private and fail closed
03

Humans curate where they already write

Shipped

The Obsidian plugin two-way syncs a local vault against Brain — local edits push, remote changes pull, and a genuine conflict parks the remote copy beside your file instead of overwriting it. The people on the team curate in the editor they already live in; the agents recall it everywhere else.

  • Two-way sync between your vault and Brain
  • Conflicts parked side by side, never silently resolved
  • Search the same knowledge base from your notes
04

Git-backed Vaults

Planned

A readable markdown mirror of each Vault, committed to a git repo you can hold — diffable, greppable, revertible, and its own exit hatch, because the export format is the repository. Designed, not shipped: it's the next thing we're building, and we'll say so here when it lands.

  • One repo per Vault, markdown you can read
  • Free tier stays user-managed — your remote or none
  • Managed remotes and tested restore are the paid line
  • Mirror a managed Vault to your own GitHub or GitLab

Today's storage is Levirge-hosted and tenant-isolated; the git mirror above is the planned durability and exit path, not a feature you can switch on yet.

Curation

Turn raw documents into cited knowledge

Plenty of products store what you give them. Brain groups incoming material by entity and produces compact summaries an agent can afford to read — each statement linking to the source documents it came from.

When a source changes, Brain rebuilds the affected summary — nobody has to remember to update it. The example below is a real one, distilled from our own August incident. Measured on a benchmark corpus, the same query returns eight fragments without that layer and one cited summary with it — the before/after.

Captured
session-capture.mddeploy-timeline.txtadr-0021-api-tokens.mdmcp-connectivity-drops.mdlistener-log.mdhandoff-thread.md
distilled
What the agent reads
entity · search-mcp

Redeploys kill in-memory MCP sessions and hour-long JWTs — clients reconnect with static per-user API tokens instead. Decided after the August connectivity incidents.

adr-0021-api-tokens.mdmcp-connectivity-drops.md+2
In the product

Shared memory, skills and handoffs for AI agents

Knowledge and curation

  • Search by meaning — semantic retrieval over everything captured, with source tracking on every fact.
  • Cited summaries — per-entity distillations that link to their documents and rebuild when a source changes.
  • Context packs — the relevant summaries, documents and relationships for an entity, assembled within a token limit.
  • Obsidian plugin — two-way sync between a local vault and Brain; search, publish and pull the same knowledge base from your notes.

Skills and handoffs

  • One skills repository — versioned skill bundles served to every connected agent as MCP tools; update one, connected agents pick it up.
  • Handoffs with context — agents send, reply and watch an inbox; work moves between agents with its context attached.
  • Delivered on reconnect — handoffs that arrive while a watcher is down are delivered on its next watch.
  • Dated reminders — park a fact against a future date — a cert expiry, a decision to revisit — and it stays out of the way until then. Brain answers "what's due in the next 30 days"; what happens next is your scheduled agent's call, not a notification we send.

Models and privacy

  • Hosted lane — hosted mode runs every model call on LLMs Levirge hosts; no third-party model provider in the path.
  • Cloud when you want speed — switch to third-party LLMs from Settings at runtime; no redeploy.
  • Local embeddings — vector generation and reranking run on Levirge infrastructure in both modes.

Access and operations

  • Tenants and Vaults — knowledge is partitioned per Vault; personal Vaults are private to their owner and fail closed.
  • Per-user tokens — hosted sign-in or bearer tokens you issue and revoke; only the hash is stored.
  • Ingestion you can see — a board of pending, errored and promoted items, with one-click retry and a health endpoint.
Teams

Personal memory is solved. This is the shared one.

Nearly every memory product remembers for one person, in one client. Brain's unit is the Vault, not the user — several people, each running their own agents, read and write the same knowledge, skills and handoffs.

The Vault is the team boundary

One Vault per project, shared by the people on it. What one person's agent learns on Tuesday, another person's agent knows on Wednesday — different client, different machine, same brain. Personal Vaults stay private to their owner and fail closed.

People and machines, same Vault

Humans sign in and curate through the web UI or Obsidian. Agents, CI jobs and long-running workers connect with per-user tokens you issue and revoke, and reach each other through handoffs — a build worker can leave a note the next session picks up.

Review, not blind trust Planned

A shared memory is only as good as your ability to audit it. Per-writer attribution on every fact, read-only and read-write scoping per token, and the git mirror that lets you diff and revert what your agents wrote are designed and being built — we'd rather name them as roadmap than imply you have them today.

Shipped today: shared Vaults, per-user tokens, handoffs between agents and CI. In build: per-writer attribution, per-token read/write scoping, git-backed review.

Where it runs

Private by default, fast when you need it

Two model lanes, one runtime switch. Hosted mode runs on LLMs Levirge hosts on its own infrastructure — no third-party model provider in the path. Cloud mode uses third-party LLMs when speed matters more than locality — and workspace data is tenant-isolated in both lanes. Self-hosting the whole system is a roadmap item, not today's offer.

Private by default

In hosted mode, every model call — categorisation, summarisation, chat — runs on LLMs Levirge hosts; no third-party model provider in the path. Flip to cloud LLMs when speed matters more than locality — a runtime setting, not a redeploy. Embeddings and reranking run on Levirge infrastructure in both modes.

We host it

Sign in and connect your clients — there's nothing to operate. Pricing is quoted per workspace and set up on the access call. Self-hosted deployment is a roadmap item for enterprise; if you need Brain to run somewhere specific, that's a conversation worth having with us.

Planned Hosted shouldn't mean held hostage. With git-backed Vaults, a Vault we manage can push-mirror to a repository you control — your GitHub, GitLab or your own server. We keep the working copy; you keep a live replica, continuously, not on request.

Private doesn't mean slow

Private inference keeps pace with ingest — categorisation, summarisation and retrieval run comfortably on a local LLM.

How it works

Every agent reads and writes one brain

Agents across every repo and client share the same memory and the same skills — contributing what they learn and drawing on what the others left behind.

Agents Levirge products Web UI users Claude Claude Desktop Codex Pi Any MCP client CI workers long-poll · CI read · write · watch Levirge Brain Knowledge Skills Handoffs Curation one workspace · agents & users share it shared context what they learn Levirge Search ✓ LIVE Your own services ✓ MCP Levirge DevOps SOON Levirge Teams SOON

One brain, one protocol — agents and users contribute to and draw on the same memory, whatever client it runs in.

Install

Pick your agent

Access is per workspace — register your interest, then connect your clients.

Early access — capacity is limited and workspaces are set up in order of request. Installing now is fine: the plugin connects the moment your workspace is active.

claude plugin marketplace add levirge/brain
claude plugin install brain@brain

Installs the /brain:* commands and connects the MCP server — sign in with OAuth on first use; workspace tokens remain as a fallback.

for agents MCP · Streamable HTTP
discovery
levirge.com/llms.txt
docs
levirge.com/brain.md
endpoint
https://brain.levirge.com/mcp
definition
levirge.com/brain/mcp.json
auth
OAuth sign-in · bearer token as fallback
verify
call kb_overview — read-only
tools
19 — brain.md for agents, docs for humans
Client config — OAuth is negotiated on connect; add an Authorization header only for token fallback.
Next step

Give your whole fleet one memory

One server, every team member. Shared knowledge that curates itself across agents and users, skills you edit once, and handoffs that carry their context.