Self-hosted MCP server + dashboard giving Claude Code, Cursor & Hermes agents shared long-term memory over a git-backed markdown vault. Agent-native, Obsidian-compatible second brain — markdown RAG, no vector DB.
TypeScript agent-memoryai-agentsanthropicclaude-codecursorgitknowledge-baseknowledge-graphllmmarkdownmcpmcp-servermodel-context-protocolnote-takingobsidianpkmragsecond-brainself-hostedzettelkasten
import path from "node:path";
/**
* Resolve a vault-relative path to an absolute one, refusing anything that escapes the vault.
*
* The write path has always had this guard. The read path did not, and `getNote` takes its
* argument straight from `brain_read` — so `brain_read({ path: "../../../../etc/passwd" })`
* read whatever the process could read. `brain_read` is a READ-scoped tool, which made this
* reachable with a read-only token: the exact token you hand an agent you don't fully trust,
* and the one the README promises is safe. On a deployed instance the reachable set included
* /proc/self/environ (AUTH_SECRET, OAuth client secrets, the Anthropic key) and the token
* hashes under ENGRAM_DATA_DIR.
*
* Every filesystem access derived from caller input goes through here. `path.resolve` collapses
* `..` before the check, so no traversal survives it, and the prefix test uses path.sep to avoid
* matching a sibling directory that merely shares a name prefix (`/vault-backup` vs `/vault`).
*/
export function resolveInVault(vaultDir: string, relPath: string): string {
const root = path.resolve(vaultDir);
const abs = path.resolve(root, relPath);
if (abs !== root && !abs.startsWith(root + path.sep)) {
throw new Error(`path escapes the vault: ${relPath}`);
}
return abs;
}
/** Non-throwing variant for read paths that already return null on a miss. */
export function tryResolveInVault(vaultDir: string, relPath: string): string | null {
try {
return resolveInVault(vaultDir, relPath);
} catch {
return null;
}
}