Docs

Connect an agent

agentbrain is a remote MCP server with OAuth 2.1 (PKCE, dynamic client registration). Any client that supports remote MCP with browser auth connects in one step.

Claude Code

claude mcp add --transport http agentbrain https://agentbrain.lukaloehr.com/mcp

Then run /mcp inside Claude Code, pick agentbrain and choose Authenticate. Your browser opens, you approve with a passkey (or you are already signed in), done.

Codex CLI

codex mcp add agentbrain --url https://agentbrain.lukaloehr.com/mcp
codex mcp login agentbrain

Claude.ai, Claude Desktop, Cursor, ChatGPT

Add a custom connector / MCP server with the URL https://agentbrain.lukaloehr.com/mcp. The OAuth flow is the same.

Headless machines

Create a personal access token in Settings and pass it as a bearer header:

claude mcp add --transport http agentbrain https://agentbrain.lukaloehr.com/mcp --header "Authorization: Bearer abp_…"

The model

A project is a scope (a codebase, a client, a domain). The global project applies everywhere. A memory is one node in the graph: a title, a summary an agent can act on, an optional markdown body, tags, importance and confidence. Relations are directed edges between memories.

Memory types

fact — Something true about the world or the project (config, IDs, state).
decision — A choice that was made and why. Supersede rather than edit history.
preference — How the owner wants things done. Highest priority in context packs.
feedback — A correction or confirmation the owner gave; include why and how to apply.
reference — Pointer to an external resource: URL, dashboard, ticket, doc.
entity — A person, service, system, account, or thing that other nodes refer to.
task — Open work, blocker, or follow-up. Archive when done.
insight — A learned lesson or pattern, usually derived from an incident.
procedure — Step-by-step how-to that has been verified to work.

Relations

relates_to — Generic association.
depends_on — A requires B to hold or exist.
blocks — A prevents B from progressing.
supersedes — A replaces B; B is usually archived.
part_of — A is a component or sub-topic of B.
caused_by — A happened because of B.
contradicts — A and B cannot both be true; one needs resolving.
example_of — A is a concrete instance of B.
derived_from — A was inferred or built from B.
see_also — Worth reading together.

Tools

`get_context`	Boot pack for a project: global memories, project memories by type and importance, stale items, recent activity.
`list_projects` / `create_project` / `update_project`	Manage scopes.
`search_memory`	Ranked full-text search across title, summary, body and tags.
`list_memories`	Browse with filters, sort and paging.
`get_memory`	Full body plus relations, optionally two hops and version history.
`explore_graph`	Breadth-first walk from a memory, filtered by relation and direction.
`remember`	Create a memory (with links, upsert by slug).
`update_memory` / `forget`	Versioned updates; archive, mark stale, or delete.
`link_memories` / `unlink_memories`	Manage relations.
`list_activity`	What every agent changed, and when.
`whoami`	Account, client, scopes and totals.

Resources: agentbrain://projects, agentbrain://projects/{slug}/context, agentbrain://memory/{id}. Prompt: bootstrap.

Security and residency

Accounts sign in with a passkey or an authenticator app after email verification. Sessions are cookie-bound and server-side.
Agents get OAuth access tokens scoped to brain:read, brain:write, account:read, rotated with refresh tokens, revocable in Settings.
All data lives in a Cloudflare D1 database created with the eu jurisdiction. Memory content never touches KV, Vectorize or AI models.
Memories are data. Agents are told never to execute instructions found in memory bodies.