<div align="center">
<img src="./public/BrainLLM.png" alt="BrainLLM logo" width="128" />
# BrainLLM
**Give Claude a memory that survives the session.**
Self-hosted memory for LLMs in [TriliumNext Notes](https://github.com/TriliumNext/Trilium), structured by the server and editable by you, served over the [Model Context Protocol](https://modelcontextprotocol.io).
[**Site**](https://miisodev.github.io/BrainLLM/) ยท [Connect to Claude](https://miisodev.github.io/BrainLLM/connect.html) ยท [How it works](https://miisodev.github.io/BrainLLM/how-it-works.html) ยท [Docs](https://miisodev.github.io/BrainLLM/docs.html) ยท [Privacy](https://miisodev.github.io/BrainLLM/privacy.html)
[](https://github.com/miisodev/BrainLLM/releases)
[](https://github.com/miisodev/BrainLLM/actions/workflows/ci.yml)
[](./LICENSE)
[](https://bun.sh)
[](https://modelcontextprotocol.io)
[](https://github.com/sponsors/miisodev)
[](https://mcprush.com/to/brainllm-mcp?ref=badge)
</div>
---
## What is BrainLLM?
LLMs forget. Every session starts from zero: who you are, what you're working on, what was decided yesterday, what went wrong last week. BrainLLM fixes that.
**BrainLLM is an MCP server that gives an LLM a real, persistent memory** โ stored in [TriliumNext Notes](https://github.com/TriliumNext/Trilium), a mature open-source knowledge base you self-host and own. The model opens each session by loading who you are and what's live, writes durable facts the moment they surface, wires knowledge together as a typed graph, and closes each session with a log, a diary entry, and a database backup. The next session picks up exactly where the last one ended.
It's a single Bun/TypeScript service with two dependencies (the MCP SDK and Zod), speaking to Trilium exclusively through its public ETAPI. Your memory lives in *your* Trilium instance โ inspectable, editable, and portable, never locked inside a vendor's black box.
### Who it's for
Anyone running an LLM as a long-running collaborator, not a one-shot Q&A tool โ across coding, research, personal knowledge management, or day-to-day operational tracking โ who wants that collaborator to actually remember: who you are, what's in flight, what was decided, without you re-explaining it every session. If you're comfortable self-hosting one small service and a Trilium instance, BrainLLM turns "the model forgot everything again" into a solved problem. It works with any MCP-capable client (Claude Desktop, Claude Code, or anything else speaking MCP), not just one product.
### From experiment to open source
BrainLLM began in February 2026 as a personal experiment: could an LLM operate a real, self-hosted memory reliably enough to be trusted as its own memory โ orienting, writing, connecting, and closing sessions without a human doing the filing? Through sustained daily, production use the answer held. The design has settled, the failure modes have been found and fixed, and the project has graduated from experiment to something **efficient and stable enough to share** โ so it's now open source. It still runs the author's own sessions every day; what you're reading is the same code, not a demo. It also still carries the shape of one person's daily use โ see [Adapting to your environment](#adapting-to-your-environment) for exactly what that means and what to change.
### The core principle
> **The model supplies content. The server owns form.**
Placement, naming, labels, deduplication, relation bookkeeping, lifecycle aging, archival, date stamping, HTML sanitization, backups โ and **structure itself** โ are all deterministic server policy, never delegated to the LLM. The model never chooses a parent note, never sets a label, never checks for duplicates, never stamps a date. That division is what makes the memory *reliable*: every guarantee is enforced at the tool layer, not requested via prompt.
### Design highlights
- **Structure is enforced, not requested** โ every content kind has a canonical structure, served by `template()` and held on write: a new thread requires its goal, thread/diary/session entries open with an identification line (which LLM, which environment, which session type), threads carry exactly one Resolution (owned by `resolve()`), duplicate section headings are detected, and `Last updated` stamps are server-maintained.
- **Domains born complete** โ creating a knowledge domain creates its book *and* its canonical Sources note (marker legend, stamp, grouped source list, revision table), so every claim has a sourcing home from the first write.
- **A visible graph** โ `graph()` renders the whole relation graph (or any note's neighborhood) as a Mermaid flowchart, maintained as a native Trilium note.
- **Byte-preserving artifacts** โ text is UTF-8, while binary content uses strict standard base64 in the tool contract and raw bytes on the wire; binary reads return an explicit MIME/base64 envelope instead of replacement characters.
- **One-call day orientation** โ `day()` serves the previous session, its change log and everything touched since in a single call.
- **Two thread shapes** โ a thread is a dated journal (one child per active day) or a titled collection (one maintained note per item, e.g. an ideas list), chosen once with `shape=` at creation.
- **Timeless notes, dated records** โ knowledge, information and thread books hold no state or history (only a domain's *Current State* does); `maintain(deep=true)` flags dated prose that creeps in. Every note except a log carries an icon, set by kind and backfilled by the sweep.
- **Resilient plumbing** โ every backend call is timeout-bounded with retry on idempotent reads; routine core writes are idempotent or duplicate-guarded, while full-mode raw overwrites remain explicit; content surgery survives the editor's own HTML rewriting; renaming a domain cascades to everything inside it; the maintenance sweep heals drift it finds.
- **Multi-agent by default** โ all write-classified tools serialize behind a process-wide FIFO lock, so two agents (an interactive session and an automated run, say) writing through one hosted instance queue in arrival order instead of racing; reads stay fully parallel. Deletion catch-up in the daily log keeps a note deleted after its day's close from vanishing without a trace โ and its window is configurable (`deletionCatchupDays` in the lifecycle policy), defaulting to Trilium's 7-day retention.
- **Ask in prose, split on a seam** โ `consistency(subject="โฆ")` finds every note asserting about a fact however it is phrased, no regex guessing; `consistency(pattern, staleAfterDays=N)` surfaces figures held in exactly one untouched note, the ones that rot silently because nothing disagrees with them, and skips sessions and diaries, which are history rather than claims; `diff(since="today")` reviews every note a session changed in one call; `revise(find="Phase 2", within="tr")` edits, inserts or removes a table row named by a few words inside it; `read(ids=[โฆ])` batches a multi-note orientation into one round trip โ each note's size up front, or a single `section=` across them; and `split(noteId, sections=[โฆ], into="โฆ")` is the write half of the oversized-note problem โ it lifts whole sections into a new note and leaves a pointer back. Information notes can carry a `#mandate` marker so a scoped session finds the one note it must obey without reading every note's prose.
- **A distinct connector identity** โ hosted instances advertise BrainLLM and their own origin, with same-origin PNG/SVG icons, explicit page favicon links and a real multi-size ICO. The canonical mark is shared across the connector, consent pages and public assets.
- **Preview changes and measure coverage** โ `revise(noteId, edits=[โฆ], dryRun=true)` shows sequential match counts and sanitized before/after content with no revision, labels or gate writes. Revision failures stop content overwrites. `consistency()` reports read failures and scan coverage; case-insensitive comparisons retain spelling variants without false conflicts. Incomplete scans cannot establish absence or unanimity.
- **Registers close in one call** โ `revise(find, within="tr", mode="remove", closure={thread, body})` removes a register row and records its closure in that thread's entry for the day, `closure={note, table?, row?}` moves a row into an explicitly selected register table; destination cells must match the table's shape. Destination-first ordering preserves recoverability, and `moves=[{find, row?}]` batches closures with preflight and per-row progress, and `append={table, row}` adds a row with no anchor to find; `within=` also accepts a row as a reader sees it (`"Phase 2 | Claude | Open"`), and a table row whose cells disagree with its header is flagged at write time and by `maintain(deep)`.
- **Bulk edits and lighter receipts** โ `revise(find, within="tr", mode="remove", all=true)` removes every row an anchor sits in, in one write and previewable with `dryRun=true`; `remember(kind="sources", revision=[{source, from}])` renames a Sources key when a source moves instead of adding a second row; a dated-thread append returns `previousRun`, its author's previous block and run number; `read(ids, sizes=true)` plans a bulk read from sizes and section sizes alone; and `diff(since="today", summaryOnly=true)` lists what changed without the lines.
- **Records you can prove** โ sessions, diary entries and dated thread entries are append-only, and now verifiably so: every `close()` seals its day's records into a digest chained to the previous day's, and `maintain(deep)` reports any sealed record that was rewritten afterwards. Appends stay legal; edits to history do not go unnoticed.
---
## How it works
At bootstrap, BrainLLM builds a five-area tree in Trilium. Every note the tools create is typed, labeled, dated, and placed by server policy:
```
BrainLLM (#brainLlmRoot)
โโโ ๐ค Master Biography ยท Goals ยท Preferences (the user โ maintained singletons)
โโโ ๐ค LLM Responsibilities ยท Protocols ยท Self-correction ยท Diary/ (the assistant's self-model + daily diary)
โโโ ๐๏ธ Memory Sessions/ ยท Threads/ (daily session logs + multi-session work)
โโโ ๐ Knowledge Master/ ยท Domains/<domain>/{ Sources, info } (learned facts beyond/contra training)
โโโ ๐ก Insights Logs/ ยท Graph ยท Claims (the brain's record of itself)
```
| Note class | Kinds | Behavior |
|---|---|---|
| **Singletons** | biography, goals, preferences, responsibilities, protocols, selfcorrection (+ each domain's Sources note) | Exactly one maintained note; edited in place; hold *current-state truth* |
| **Authored records** | diary, session, dated threadEntry | One per calendar day and parent; every write lands as a timestamped addendum block โ chronology is the point |
| **Generated logs** | log | One per calendar day; regenerated in place from the change feed |
| **Collections** | thread, user, information, domain | Titled notes, deduplicated by normalized title within their scope |
A session follows an enforced protocol: `start()` orients (who you are, live threads, what changed since last time) โ the model works, writing durable facts *as they surface* โ `session()` โ `addendum()` โ `maintain()` โ `remarks()` โ `diary()` โ `close()` commits the log, regenerates the daily change log, and triggers a DB backup. The pre-close gate is **enforced in code**: `close()` refuses until the required steps actually ran and the last `session()`, `remarks()` and `diary()` calls occurred in that order โ narrating "I did the steps" doesn't count, only tool calls do. The gate is *durable*, written to the session note rather than held in memory, so it survives a restart mid-session and behaves the same on stdio and HTTP. A session that wrote no content gets a **light close**: nothing to fold or sweep, so `close()` asks only for `session()`, `remarks()` and a short `diary()`.
`start()` returns Preferences and Protocols **in full**, so the model has the operating rules immediately. Biography, Goals, Responsibilities and Self-correction arrive as **section headings and previews**; pull the sections the work needs. `start(depth="full")` inlines all of them.
Knowledge is a **typed graph**: a closed vocabulary of 16 relations (`extends`, `contradicts`, `supports`, `partOf`, `supersedes`, `corrects`, โฆ), wired by `connect()` or at creation, traversed by `explore()` (links / backlinks / neighborhood / shortest path), rendered by `graph()`, and audited by `maintain(deep)`, which heals duplicate edges and flags what a note application would never think to look for: orphaned notes, stale content, duplicate titles, structural drift *inside* a note, bodies past the read ceiling, stubs that were labelled and never written, titles carrying a date (which defeats dedup-by-title), threads heavy enough to want consolidating, Sources notes whose verification table was never filled in, and bodies carrying doubly-escaped markup. Most of it is computed from note properties server-side rather than by reading bodies, so the report is complete rather than truncated. Findings can be acknowledged (`maintain(ack=[โฆ])`) so a flag you have judged correct goes quiet until that note's content actually changes โ a warning that reappears every run and is correctly ignored every run trains you to skim the list.
---
## Quick start
### Prerequisites
- [Bun](https://bun.sh) v1.4.2+
- A running [TriliumNext](https://github.com/TriliumNext/Trilium) instance (desktop app or server)
### 1. Install and build
```bash
git clone https://github.com/miisodev/BrainLLM
cd BrainLLM
bun install
bun run build
```
### 2. Get an ETAPI token
**Or skip this step.** Set `TRILIUM_PASSWORD` instead of `TRILIUM_ETAPI_TOKEN` and BrainLLM mints a token itself on first start, caching it beside `brainllm.json` so restarts reuse it rather than creating a new one each time. This is the easiest path for a container deploy, where there's no terminal to run a setup command in.
To do it explicitly instead, either let `init` mint one:
```bash
TRILIUM_BASE_URL=http://localhost:8080 TRILIUM_PASSWORD=your-trilium-password bun run init
```
It caches the token beside `brainllm.json` with owner-only permissions and does **not** print the secret. If you explicitly need a copy, run `bun run init -- --show-token`; the CLI is idempotent and adopts an existing `#brainLlmRoot` instead of creating a duplicate tree. Or create one by hand in Trilium: **Options โ ETAPI โ Create token**.
### 3. Configure your MCP client
For **Claude Desktop**, add to `claude_desktop_config.json` (see `config.example.json` for a complete example including remote):
| Platform | Path |
|----------|------|
| Windows | `%APPDATA%\Claude\claude_desktop_config.json` |
| macOS | `~/Library/Application Support/Claude/claude_desktop_config.json` |
| Linux | `~/.config/Claude/claude_desktop_config.json` |
```json
{
"mcpServers": {
"BrainLLM": {
"command": "bun",
"args": ["run", "/absolute/path/to/BrainLLM/dist/index.js"],
"env": {
"TRILIUM_BASE_URL": "http://localhost:8080",
"TRILIUM_ETAPI_TOKEN": "your-etapi-token-here"
}
}
}
}
```
For **Claude Code**, the same server block works in `.mcp.json`. A schema-valid MCPB manifest (`manifest.json`) is included for hosts that support custom Bun commands, but BrainLLM still requires Bun and this repository does not yet publish a `.mcpb` one-click artifact.
### 4. Bootstrap the brain
Restart your client and ask Claude to run `bootstrap` โ it builds the five-area tree (each area engraved with its purpose), writes `brainllm.json`, and is active immediately. Or from the CLI:
```bash
TRILIUM_BASE_URL=http://localhost:8080 TRILIUM_ETAPI_TOKEN=your-token bun run init
```
### 5. Teach the model to operate it
Install the **operational skill** โ the document that teaches Claude to treat BrainLLM as its own mind rather than a filing cabinet:
```bash
# Claude Code / Cowork skills directory (macOS/Linux)
cp -r skills/brainllm ~/.claude/skills/brainllm
```
```powershell
# Windows PowerShell
New-Item -ItemType Directory -Force "$env:USERPROFILE\.agents\skills" | Out-Null
Copy-Item skills\brainllm "$env:USERPROFILE\.agents\skills\brainllm" -Recurse -Force
```
For clients that take a skill upload (claude.ai), use `brainllm.zip` from the [latest release](https://github.com/miisodev/BrainLLM/releases/latest): it is the same `skills/brainllm/` folder, zipped.
Then start a session. The model calls `start()`, orients, and operates the brain natively from there.
---
## The tool surface
All tool returns are structured JSON. User-input mistakes return informational errors (`{error, detail, hint}`) the model can read and self-correct from; core note reads, writes and searches include a free `relations` snippet; routine core writes use duplicate guards or upserts. A lost response can follow an accepted write; inspect the receipt and affected notes before retrying, especially raw full-mode mutations.
### Why the tools are named the way they are
**The naming is a readability feature for you, not a style choice.** Most MCP clients show you the name of each tool call as it happens. BrainLLM's tools are named as **plain-English verbs for what is happening to your memory** โ `remember`, `recall`, `revise`, `resolve`, `forget`, `connect`, `maintain` โ so that the call log reads as a description of what the model is doing in your brain, without you having to know the schema or open Trilium to find out.
The alternative, which most servers pick, is to name tools after the API underneath: `create_note`, `patch_note`, `update_attribute`. That tells you a row changed. It doesn't tell you the model recorded a decision, corrected something it previously believed, or closed a line of work โ and those are the things you actually want to see scroll past.
Read this way, a session is legible at a glance:
| You see | What it means |
|---|---|
| `start` โ `day` | Orienting: loading who you are, what's open, what changed since last time |
| `remember` ยท `diary` | Writing something down โ a durable fact, or the model's own daily record |
| `recall` ยท `domain` ยท `outline` | Looking something up before answering, rather than answering from training |
| `revise` ยท `resolve` ยท `forget` | Changing, completing, or archiving something that already existed |
| `connect` ยท `graph` | Wiring a relationship between two things it noticed |
| `maintain` ยท `addendum` | Housekeeping โ aging, structural drift, folding pending edits in |
| `remarks` โ `close` | Wrapping up: reflecting, then committing the session log |
Three conventions keep that legible as the surface grows:
- **Verbs for actions, nouns for surfaces.** `remember` writes; `knowledge` reads. The five surface nouns (`master`, `llm`, `memory`, `knowledge`, `insights`) only read; names such as `graph` and `session` also write.
- **`<surface>` / `<surface>_recall` pairs.** The bare name reads a surface in full; the `_recall` suffix skims it. Seeing `memory_recall` rather than `memory` tells you the model chose the cheap read.
- **Destructive verbs say so.** `forget`, `detach`, `delete_*` are never spelled as something softer, so an irreversible call is never disguised as a routine one.
The **full mode** tools deliberately break this convention โ they are named after the raw Trilium primitives (`create_note`, `patch_note`, `delete_attribute`) precisely so that a raw, guard-free operation is visibly different in the log from a brain-aware one.
### Tool permissions
Every tool declares whether it reads, writes, or destroys, so your client can group them and you can grant permission per group instead of per call. Every tool also carries a human-readable title ("Edit a note", "Search the brain") for permission prompts and directory listings. The current full-mode split is **34 read-only ยท 31 write ยท 13 destructive** (core mode is **23 read-only ยท 17 write ยท 5 destructive**); `revise`, `label` and `connect` count as destructive because they can overwrite or remove what was there, which lets reads run unattended while anything that touches your brain still asks.
The classification lives in one reviewable table ([`src/annotations.ts`](./src/annotations.ts)) rather than scattered across registrations, and its default is deliberately unsafe-side: a tool missing from the table is treated as a **write**, never a read, and says so at startup. Three are worth knowing because they look like reads and aren't โ `start()` creates today's diary and session stubs, `session()` runs the maintenance sweep, and `graph()` writes the rendered graph note.
### Core โ universal verbs (35)
| Group | Tools |
|---|---|
| Session lifecycle | `start` ยท `day` ยท `session` ยท `remarks` ยท `close` ยท `backup` ยท `health` |
| Writing | `remember` ยท `diary` ยท `revise` ยท `split` ยท `resolve` ยท `withdraw` ยท `recover` |
| Reading & search | `read` ยท `recall` ยท `domain` ยท `brain` ยท `assembly` ยท `outline` ยท `inspect` ยท `diff` ยท `template` ยท `consistency` |
| Graph | `connect` ยท `explore` ยท `graph` |
| Attachments & labels | `attach` ยท `detach` ยท `label` |
| Maintenance & system | `addendum` ยท `maintain` ยท `claim` ยท `forget` ยท `bootstrap` |
### Core โ surface reads (10, dual-mode)
`master`/`master_recall` ยท `llm`/`llm_recall` ยท `memory`/`memory_recall` ยท `knowledge`/`knowledge_recall` ยท `insights`/`insights_recall` โ each surface read in full, or skimmed as compact stubs.
### Full mode (`BRAINLLM_MODE=full`, +33)
The raw ETAPI, one tool per Trilium primitive โ notes, attributes, branches, revisions, attachments, calendar, system. Outside core placement, sanitising and revision policy: an edge-case fallback for what core can't express (precise raw queries, code/canvas/mermaid notes, branch surgery, revision recovery).
The complete operational reference is [`skills/brainllm/SKILL.md`](./skills/brainllm/SKILL.md); per-topic deep dives live in [`skills/brainllm/references/`](./skills/brainllm/references/). Beyond that, the **architecture reference** explains how BrainLLM is built and why โ every module, every tool's behaviour, the sweep internals, the idempotency matrix โ and is kept current with each release rather than published in this repo; see [Support the project](#support-the-project) for how to get a copy.
---
## Deployment
### Transport modes
| Mode | When | Selected by |
|------|------|-------------|
| **stdio** | Local โ Claude Desktop / Claude Code spawns BrainLLM as a child process | `PORT` unset (default) |
| **HTTP connector** | Remote โ clients reach BrainLLM over the network | `PORT` set (Railway injects it) |
The HTTP connector serves a streamable-HTTP MCP endpoint at `/mcp` (one session per `mcp-session-id`, DELETE terminates, CORS-enabled with `mcp-session-id` exposed for browser clients), a **legacy SSE endpoint** at `/sse` (with `POST /messages`) for clients that predate streamable HTTP โ older Cursor builds, Continue, and Python-SDK clients โ and both `GET /health` and `GET /ready`. The server root serves a plain page for ordinary browser `GET` requests; MCP-shaped requests sent to `/` are also accepted as a compatibility alias, because some hosted Claude connector versions normalize the post-token MCP path to the origin root. The server itself speaks plain HTTP โ TLS is expected to terminate in front of it (Railway's edge does this automatically for you; on a bare VPS or your own Docker host, put it behind a reverse proxy such as Caddy, nginx, or a Cloudflare Tunnel). **HTTP mode now fails closed unless `MCP_AUTH_TOKEN` or `BRAINLLM_OWNER_PASSWORD` is configured**; `BRAINLLM_ALLOW_UNAUTHENTICATED_HTTP=true` is an explicit escape hatch for an isolated trusted network only. CORS defaults to `*`, so a reachable deployment must have a real authentication gate. Idle sessions are evicted after 1 hour; MCP request bodies are capped at 50 MiB by both the server and SDK transport. Both transports sit behind the same authentication gate.
### Health, readiness and shutdown
`GET /health` is liveness only: it returns `OK` once the HTTP listener can answer and never contacts Trilium. `GET /ready` is readiness: it calls Trilium's `app-info` (the cheapest authenticated read that proves reachability and token validity) and returns `{"status":"ready"}` with 200, or `{"status":"unavailable"}` or `{"status":"draining"}` with 503. A healthy result is cached for 10 seconds and a failure for 5; each check is bounded at 3 seconds, concurrent probes share one backend call, and a check that outlives its deadline is not abandoned and re-issued. The response carries only the state, never the Trilium URL, token, response body or error.
Both endpoints are unauthenticated and excluded from request logging. In HTTP mode, `SIGTERM` and `SIGINT` mark the process draining: `/ready` reports `draining`, new work on every other route gets 503 with `Retry-After: 5`, and the listener stays open while in-flight requests finish. The whole drain is bounded at 20 seconds; transports and the HTTP server are then closed within the remaining grace and the process exits 0. Stdio installs attach no signal handler and keep their EOF behaviour.
HTTP requests, MCP calls and shutdown phases are JSON lines on stderr (`[brainllm] {...}`) carrying `ts`, `event`, `release` and only allowlisted scalar metadata: HTTP method, fixed route label, tool name, status, duration, shutdown phase/signal and in-flight count. Request bodies, tool arguments and results, JSON-RPC error text, URLs, note ids and tokens never enter a log line; unregistered methods and tools are recorded as `unknown`, and `/health`, `/ready` and asset requests are not logged.
### Docker / Railway
The included [`Dockerfile`](./Dockerfile) builds and runs the HTTP connector (two-stage, digest-pinned `oven/bun`, drops privileges via `entrypoint.sh`). On **Railway**: `PORT` is auto-injected โ set `MCP_AUTH_TOKEN`, `TRILIUM_BASE_URL`, `TRILIUM_ETAPI_TOKEN`, and `BRAINLLM_TZ` as service variables, deploy, and point your client at `https://<app>.up.railway.app/mcp` with the bearer token. Deploying elsewhere (a VPS, your own Docker host) works the same way โ the container only needs those same env vars and a TLS-terminating proxy in front of it, per the transport note above.
Unexpected HTTP exceptions, internal JSON-RPC errors and failed tool results hold `/ready` unavailable for 30 minutes, or until restart, so an external probe can alert even when Trilium is reachable. Ordinary authentication and protocol rejections do not trigger this signal. `/health` stays live and accepted MCP work remains available; inspect the safe error metadata before restarting.
Browser MCP calls must carry an allowed `Origin`: the configured `BRAINLLM_PUBLIC_URL` origin, literal loopback origins at the listening port, or an exact entry in comma-separated `BRAINLLM_ALLOWED_ORIGINS`. Configure an Inspector/browser integration explicitly; wildcard, opaque `null` and malformed origins are refused with 403, including preflights. Native clients without `Origin` are unaffected. The server never builds this allowlist from an incoming Host header.
### Branding
In HTTP mode BrainLLM serves its own icons โ `/icon.png`, `/icon.svg`, sized variants, and `/favicon.ico` โ and advertises them in the MCP handshake together with its title, a one-line description and its website, so clients show the BrainLLM mark in your connector list and beside its tool calls. The handshake also carries short server instructions describing the five areas and the session shape. Everything brand-shaped (name, tagline, description, the mark's geometry) is defined once in [`src/brand.ts`](./src/brand.ts), and a unit test holds `package.json`, `manifest.json` and `server.json` to the same description.
Both halves of that are load-bearing. The spec tells clients to **verify icon URIs share the server's origin**, so icons hosted anywhere else are silently dropped. And a host that serves no favicon makes clients fall back to the *registrable domain's* โ a BrainLLM at `brain.yourdomain.com` would show `yourdomain.com`'s site logo, which is worse than nothing because it looks intentional.
To use your own artwork, replace `public/BrainLLM.svg` and run `bun run icons` โ it rasterises the vector at each size the server serves. Commit the results; the Docker image copies `public/` and has no image toolchain at runtime.
### Connecting remote clients (OAuth)
Claude Code and `mcp-remote` send `MCP_AUTH_TOKEN` as a header and need nothing further. **The hosted Claude surfaces โ claude.ai, Claude mobile, Cowork โ cannot.** Their custom-connector UI offers OAuth or nothing; there is no field for a bearer token. So BrainLLM ships its own OAuth 2.1 authorization server. It serves **both** registration mechanisms, so every OAuth-capable client works:
- **Client ID Metadata Documents (CIMD)** โ the registration-free mechanism adopted in the MCP `2025-11-25` specification; this is what Claude selects.
- **Dynamic Client Registration (RFC 7591)** โ for clients that never grew CIMD support. opencode (MCP TS SDK โค1.29) is the forcing case: without a `registration_endpoint` it refuses to authenticate at all.
Set one variable to turn it on:
```bash
BRAINLLM_OWNER_PASSWORD=<a strong password you choose>
```
That password is what you type on the consent screen. Without it the OAuth endpoints are disabled entirely and nothing else changes.
#### Optional owner second factor (TOTP)
Set `BRAINLLM_OWNER_TOTP_SECRET` as well to require a six-digit authenticator code alongside the owner password. It is RFC 6238 TOTP: HMAC-SHA1, six digits, 30-second steps, with one step of clock tolerance either way. The secret must be base32 and at least 128 bits (16 bytes); lowercase, spaces and hyphens are accepted. While it is set, `BRAINLLM_OWNER_PASSWORD` must be at least 16 characters. A configured but invalid or too-short secret, or an MFA password under 16 characters, fails closed with a 503 "misconfigured" page โ it never silently falls back to password-only.
**Enroll before enabling.** Add the secret to an authenticator app as a time-based, six-digit, 30-second entry, confirm it produces a current code, and keep an offline backup of the seed. Then set the variable, redeploy, and complete one consent sign-in to confirm the code field appears. Enabling the variable before the authenticator holds the seed locks you out of new approvals.
A used code cannot be replayed. The last accepted step is persisted in the OAuth store beside `brainllm.json` with a SHA-256 fingerprint of the seed, never the seed itself; the store survives a process restart, and on a persistent-volume deployment a redeploy too. The second factor gates consent approval only: already-issued access and refresh tokens, and `MCP_AUTH_TOKEN`, keep working when it is enabled or removed.
Recovery is deliberately administrative. If the authenticator is lost, restore the seed from the offline backup. If there is no backup, a deployment admin removes `BRAINLLM_OWNER_TOTP_SECRET` and redeploys, returning that deployment to password-only; there is no in-app bypass.
Then in Claude: **Settings โ Connectors โ Add custom connector**, enter `https://<your-host>/mcp`, and leave the OAuth client fields blank โ CIMD needs no client ID or secret. Claude discovers everything from the `401`, you approve once on a consent screen, and it stays connected.
What that turns on:
| Endpoint | Purpose |
|---|---|
| `/.well-known/oauth-protected-resource` (+ `/mcp` variant) | RFC 9728 โ names the authorization server |
| `/.well-known/oauth-authorization-server` (+ `/mcp` variant) | RFC 8414 โ advertises CIMD support, DCR, and PKCE S256 |
| `/authorize` | Validates the client (CIMD document or registered id), shows the consent screen |
| `/token` | PKCE-verified code exchange, with rotating refresh tokens stored only as SHA-256 hashes |
| `/register` | RFC 7591 dynamic client registration โ how opencode and other DCR-only clients connect |
The consent screen names the requesting client and the address it will return to, and warns when that address is your own computer (a loopback redirect, as Claude Code uses). MCP traffic is rate-limited per bearer token, with a looser per-IP bound, so clients that share an egress range (all of Claude's hosted traffic does) never exhaust each other's budget.
Both credentials work simultaneously โ a static token from Claude Code and an OAuth token from claude.ai, against the same brain. Access tokens are signed JWTs bound to your server's resource URI, so a token minted for a different MCP server is rejected; the signing secret and refresh tokens persist beside `brainllm.json`, which on a container deploy means **putting `BRAINLLM_CONFIG` on a volume** โ otherwise every redeploy invalidates every token.
For config persistence across redeploys, mount a volume on the **BrainLLM service** (not the Trilium service) and set `BRAINLLM_CONFIG` to a file path inside it:
```bash
railway volume -p <project-id> -s <brainllm-service-id> -e production add --mount-path /vol
railway variables set BRAINLLM_CONFIG=/vol/brainllm.json
```
Without a volume, leave `BRAINLLM_CONFIG` unset โ auto-discovery re-finds the brain from Trilium's `#brainLlmRoot` marker on each cold start (~1 s).
### Configuration reference
| Variable | Required | Purpose |
|---|---|---|
| `TRILIUM_BASE_URL` | โ
| URL of the TriliumNext instance (local, or an HTTPS reverse-proxy/tunnel URL) |
| `TRILIUM_ETAPI_TOKEN` | โ | ETAPI bearer token. Required unless `TRILIUM_PASSWORD` is set |
| `TRILIUM_PASSWORD` | โ | Alternative to the token: BrainLLM mints one on first start and caches it beside `brainllm.json`, reusing it across restarts. Also used by `bun run init` |
| `BRAINLLM_MODE` | โ | `core` (default) or `full` (adds the raw ETAPI surface) |
| `BRAINLLM_TZ` | โ | IANA timezone (e.g. `Africa/Johannesburg`) so dates stamp in *your* day on hosted deploys; unset = host clock |
| `BRAINLLM_DELETION_CATCHUP_DAYS` | โ | How many days your Trilium keeps deleted notes (Options โ Other โ "Erase notes after"). Deletions are reported only within this window, and ETAPI cannot read the option. Default `7`, Trilium's own default; overrides `policy.deletionCatchupDays` |
| `PORT` | โ | Presence switches to HTTP-connector mode |
| `MCP_AUTH_TOKEN` | โ | Static bearer token accepted on `/mcp` โ what Claude Code and `mcp-remote` send |
| `BRAINLLM_OWNER_PASSWORD` | โ | Enables the OAuth 2.1 / CIMD endpoints, required for claude.ai. The password you type on the consent screen |
| `BRAINLLM_ALLOWED_ORIGINS` | No | Comma-separated exact HTTP(S) origins for browser MCP clients, in addition to the configured public origin and literal loopback origins at PORT. No wildcard, path, userinfo or opaque `null` |
| `BRAINLLM_OWNER_TOTP_SECRET` | โ | Optional RFC 6238 owner second factor: base32 seed of 128+ bits, six-digit/30-second codes. Requires `BRAINLLM_OWNER_PASSWORD` of 16+ characters; enroll before enabling |
| `BRAINLLM_OAUTH_SECRET` | โ | Optional 32+ character deployment signing secret; rotate it to invalidate old OAuth tokens |
| `BRAINLLM_ALLOW_UNAUTHENTICATED_HTTP` | `false` | Explicit trusted-network-only escape hatch when no HTTP credential is configured; never use on a public interface |
| `BRAINLLM_TRUST_PROXY` | `false` | Trust `x-forwarded-host`/`x-forwarded-proto` for OAuth discovery only when a trusted reverse proxy sets them; prefer `BRAINLLM_PUBLIC_URL` |
| `BRAINLLM_PUBLIC_URL` | โ | Override the derived public origin when a proxy rewrites `Host`. Must match the URL you enter in the client exactly |
| `BRAINLLM_CONFIG` | โ | Absolute file path for `brainllm.json` on persistent-volume deploys |
Lifecycle timings live in `brainllm.json` and are yours to tune:
```json
"policy": { "dormantAfterDays": 21, "archiveDormantAfterDays": 45, "staleAfterDays": 7 }
```
Threads age `active โ dormant โ archived-in-place` on that timeline; nothing is ever deleted by aging, and `recover()` / `withdraw()` bring anything back.
---
## Adapting to your environment
A handful of values in this repo reflect the author's own machine. None are secrets, but if you clone or fork, review these and swap in your own:
| What | Where | Make it yours |
|---|---|---|
| **Timezone** | `BRAINLLM_TZ` in `.env` | Your IANA zone โ or unset for the host clock |
| **Deleted-note retention** | `BRAINLLM_DELETION_CATCHUP_DAYS` in `.env` | Only if your Trilium erases deleted notes after something other than 7 days |
| **Config path** | `BRAINLLM_CONFIG` env var | Only needed on persistent-volume deploys (see above) |
| **Bundle path** | your MCP client config | The real absolute path to `dist/index.js` on your machine |
| **Author ยท repo ยท funding** | `package.json`, `.github/FUNDING.yml`, the badges above | Your own details if you fork; the funding links support the original author |
---
## Architecture
```
index.ts โโฌโ tools.ts โโฌโ trilium.ts ETAPI client: bounded/retrying I/O, graph traversal
โ โโ router.ts placement policy: kind โ parent, label plans, domains
โ โโ templates.ts canonical structures per kind + the template-tool rules
โ โโ normalize.ts titles, slugs, markdownโHTML, sanitization, surgery
โ โโ elements.ts the element (row, item, paragraph) an anchor sits in
โ โโ diffing.ts block and table-cell diffs, revision selection
โ โโ lifecycle.ts protection, sweep, aging, the start digest
โ โโ journal.ts daily Insights log generation (regenerate-in-place)
โ โโ seal.ts per-day record seals that maintain(deep) verifies
โ โโ bootstrap.ts five-area tree builder
โ โโ tools-{master,llm,memory,knowledge,insights}.ts โ tools-surface.ts
โโ tools-advanced.ts full-mode raw ETAPI
โโ serialize.ts ยท coerce.ts ยท annotations.ts wrap every tool: write lock, param coercion, hints and titles
โโ cap.ts ยท brand.ts ยท config.ts the 140k result cap ยท name, mark, instructions ยท brainllm.json
โโ HTTP: sse.ts ยท rate-limit.ts ยท operability.ts legacy SSE ยท public budgets ยท readiness, logs, drain
โโ OAuth: oauth.ts ยท totp.ts OAuth 2.1 (CIMD checks in network-security.ts) ยท optional owner TOTP
```
Shared modules: `types.ts` defines vocabularies, `time.ts` handles local dates, and `icons.ts` supplies note icons. The separate `init.ts` CLI bootstraps and saves configuration; the server entry point is `index.ts`. The tree and shared modules account for all 35 runtime modules under `src/`; the `*.test.ts` suites and the `test.ts` live harness are separate.
Key properties: no tool result exceeds 140,000 characters (a larger one is cut with a note naming the narrower read); core text content is sanitized for Trilium/CKEditor 5 compatibility (mutations reported back as `sanitized[]`); a revision snapshot precedes routine content mutation; every request carries the user's local time so Trilium stamps dates in the right day; core writes are idempotent or duplicate-guarded, while raw full-mode overwrites are explicit.
## Development
```bash
bun run dev # hot-reload dev server
bun run build # bundle to dist/index.js
bunx --bun tsc --noEmit # typecheck (the bundler does not)
bun run test:unit # all unit tests (Bun discovery; works on Windows and CI)
bun run test:http # local auth/header smoke (no Trilium required)
BRAINLLM_LIVE_TOKEN=... bun run test:live # read-only smoke against a deployed server, over real MCP
bun run test # integration tests (requires a live Trilium instance)
bun run validate:manifest # MCPB schema/icon validation
bun run validate:release # every version site, server.json schema, SECURITY.md line
bun run init # CLI bootstrap
```
Repo layout: runtime source in `src/`, the operational skill in `skills/brainllm/`, developer scripts in `scripts/`, schema-valid MCPB manifest in `manifest.json`, container build in `Dockerfile`.
## Contributing
Contributions are welcome โ bug reports, fixes, docs, and ideas alike.
### What we value
BrainLLM started as one person's daily-driver tool, so the code and docs still carry that person's assumptions in a few places (see [Adapting to your environment](#adapting-to-your-environment)). The contributions that matter most are the ones that generalize it beyond that:
| Value | What that looks like in a PR |
|---|---|
| **Generalization** | Replacing an assumption baked in for one workflow with something that holds for any brain's content and conventions |
| **Cross-user** | Nothing in `src/` should assume a specific person's ventures, schedule, or vocabulary โ that belongs in *content* the model writes, never in the tool layer |
| **Cross-device** | Behaves the same on Windows/macOS/Linux, over local stdio and remote HTTP, on a laptop and in a Railway container |
| **Token efficiency** | Smaller responses, fewer round-trips, cheaper reads โ e.g. `recall()`'s ranked snippets over `inspect(content=true)` for a plain yes/no check. This is a standing design bias, not a one-off optimization |
| **Latency** | Bounded, retrying I/O; parallel reads over sequential ones; nothing that makes session start or a tool call slower than it needs to be |
### How to contribute
1. **Open an issue first** for anything non-trivial โ bugs with reproduction steps, or proposals with the use case spelled out.
2. **Fork and branch**, keep changes focused, and match the existing code style.
3. **`bun run build` must pass clean** and unit tests must stay green (`bun run test:unit` โ CI runs the same checks on every push and PR); add tests for new normalize/lifecycle logic.
4. **Use conventional commits** (`fix:`, `feat:`, `docs:`, `refactor:`).
5. **Update the docs that your change touches** โ `README.md`, the skill package under `skills/brainllm/`, and `config.example.json`/`.env.example` where relevant. The skill is part of the product: a tool change without its skill update is half a change.
Not sure where to start? Issues labeled `good first issue`, doc gaps, and the troubleshooting reference are all friendly entry points.
## Security
Please do not report suspected vulnerabilities in a public issue. Follow the private reporting process in [`SECURITY.md`](./SECURITY.md); never include ETAPI tokens, OAuth credentials, or private brain content in a report.
If the OAuth store is unreadable or its replay state is corrupt, new MFA grants fail closed across restarts. Preserve the damaged file for diagnosis and restore a verified OAuth-store backup, or deliberately reset it through deployment administration after rotating the OAuth signing secret and arranging client reconnection. Do not repeatedly delete state to work around an unexplained error.
### Verify a package
Install the exact version in a disposable directory with `npm install --ignore-scripts brainllm@<version>`, then run `npm audit signatures`. This verifies registry signatures and available provenance attestations; it does not prove the code is safe. Open the package's Provenance link (also discoverable with `npm view brainllm@<version> dist.attestations --json`) and check the expected source repository `miisodev/BrainLLM`, workflow `.github/workflows/publish.yml`, tag `refs/tags/v<version>` and that tag's commit. Reject unexpected identities or verification failures. See [npm's provenance guidance](https://docs.npmjs.com/generating-provenance-statements/).
## Privacy and support
BrainLLM adds no analytics, telemetry or hosted account. Notes are stored in your Trilium, and sign-in records stay on your server. HTTP mode's structured request, MCP and shutdown logs are metadata only: no note content, tool arguments or results, URLs, note ids or tokens. Connected MCP clients receive the content their tools read, under those clients' own data policies. The full [privacy policy](https://miisodev.github.io/BrainLLM/privacy.html) says what is stored where. For help, open a [GitHub issue](https://github.com/miisodev/BrainLLM/issues) or email miisodev@gmail.com.
---
## Support the project
BrainLLM is built and maintained by one person. If it's useful to you, sponsorship directly funds its continued development:
<div align="center">
[](https://github.com/sponsors/miisodev)
[](https://paypal.me/miisodev?locale.x=en_US&country.x=ZA)
</div>
Sponsors can request a copy of the **architecture reference** โ the internal guide to the modules, tool behavior, content pipeline, lifecycle and deployment that isn't published in this repo. Message via GitHub Sponsors to ask for one.
Starring the repo, reporting bugs, and spreading the word help too.
## License
[MIT](./LICENSE) ยฉ [Kevin Miiso Novo](https://github.com/miisodev)
## Credits
- [TriliumNext Notes](https://github.com/TriliumNext/Trilium) โ the open-source, self-hosted knowledge base that powers this server's backend. BrainLLM would not exist without the TriliumNext team's work.
- [Model Context Protocol](https://modelcontextprotocol.io) โ the open standard that lets one memory serve every MCP-capable client.