Alexandria

🏛️ Alexandria

> A local knowledge vault for your AI agents — saves tokens, remembers across sessions, and turns every model into a disciplined engineer.

[](https://www.npmjs.com/package/@ureck/alexandria) [](./LICENSE) [](https://buymeacoffee.com/ureck)

[](https://ureckchan.github.io/alexandria/) Documentation / Documentación

Alexandria automatically captures your AI prompts and sessions, searches that knowledge before spending tokens, and connects everything in an Obsidian-style graph. 100% local: no database, no API keys, no telemetry, no cost.

Works with the most popular agents via MCP: Claude Code · Cursor · OpenCode · Devin CLI · Cline · Codex CLI · Gemini CLI · VS Code (Copilot) · OpenClaw · Hermes Agent · Windsurf (legacy).

Why?

Every new AI session starts from zero: it re-explores your code, re-derives decisions you already made, re-spends the same tokens. Alexandria breaks that cycle:

• Saves tokens: every prompt triggers a local semantic search ( Plan → Execute one task → Verify with evidence → (on failure: reuse past lessons) → Extract the lesson

It lives in the MCP tools themselves — their names and descriptions carry the instructions — so it works identically across every agent, with or without hooks:

| MCP tool | What the agent does with it | Token payoff | |---|---|---| | plan_create(title, goal, dod, tasks, design?) | Plans like an architect: clarifies vague goals first, optionally records the architecture (design → a ## Design section), and the response surfaces similar past plans + relevant lessons plus quality hints (verifiable DoD, small tasks) — suggestions, never a gate | No wandering; open plans are re-injected next session so work is resumed, not repeated | | task_verify(plan, task, passed, verify_command, cwd?) | Runs verify_command (e.g. npm test) and takes the verdict from the real exit code — not from the agent's word. Flags a discrepancy if the agent claimed otherwise; without a command the note is tagged self-reported and ale audit calls it out | "✅ 12/12 passed" (5 tokens) instead of a 500-line log; on failure it automatically surfaces past lessons that match | | lesson_extract(title, lesson, links) | Distills the root cause + fix after solving something non-obvious | ~200 tokens today replace ~20,000 tokens of re-debugging next month |

Lessons and solutions get a ranking boost in search, so the most valuable knowledge always surfaces first. The protocol is on by default; toggle it anytime:

ale config set protocol false   # classic vault only
ale config set protocol true    # back on
ale init --no-protocol          # install without it

Install

# npm (Node ≥ 20)
npm install -g @ureck/alexandria

# bun
bun add -g @ureck/alexandria

# deno
deno install -grA -n ale npm:@ureck/alexandria/ale
# (if Deno complains about "minimum dependency date", add --minimum-dependency-age=0)

Verify:

$ ale --version
1.3.1

Getting started

One command installs everything (embedding model ~130MB downloaded once, hooks, MCP registration, index):

# Global — one shared vault for all your sessions (~/Alexandria)
ale init

# Per project — vault INSIDE the repo (./Alexandria)
cd my-project
ale init --project

# Point at an existing Obsidian vault
ale init --path ~/Documents/MyObsidianVault

# Choose which agents get the MCP server
ale init --agents all                      # every supported agent
ale init --agents claude,cursor,opencode   # just these
# (without --agents: Claude Code + whatever is detected on your machine)

ale init --project also, automatically:

• adds the vault to your .gitignore (personal knowledge, not repo code),
• scans the project (stack from package.json, npm scripts, folder structure, README excerpt) into an initial Map - note — context from the very first session,
• asks "Configure Claude to use Alexandria automatically?" and writes the usage rules into CLAUDE.md (--auto / --manual to skip the question).

After init there is nothing to reconnect, ever. Important: hooks load when a session starts — open a new agent session in that folder (an already-open one won't see them).

Daily usage — examples

Search by meaning, not keywords

$ ale search how do I publish my app to the google store

Flutter release deployment (Alexandria/notes/2026-07-09-deploy.md, note) ● high relevance
  To publish on Play Store: keystore outside git, key.properties,
  flutter build appbundle. Always sign with the production key.

The query shares zero keywords with the note ("google store" → "Play Store"). More:

ale search cors error in the api        # finds your fix from months ago
ale search "architecture decision" -k 10
ale search auth --expand                # also pulls notes CONNECTED to the results (graph)

The ● high/medium/low relevance label is relative to the best hit — trust it over the raw number.

Save notes manually

# Quick note
ale add "Prisma setup" -c "Singleton in src/lib/prisma.ts using globalThis" -t "prisma,db"

# From a file in your project (or a pipe)
ale add "Q3 architecture decisions" --file DECISIONS.md
cat DECISIONS.md | ale add "Q3 architecture decisions"

# With [[wikilinks]] to connect knowledge
ale add "Vercel deploy" -c "Region cdg1, see [[Prisma setup]]" -t deploy

(Automatic capture via hooks already does this in every Claude Code session — add is for extra knowledge.)

Control what spends tokens (ale toggles)

Every token-spending behavior is opt-out — before or after installing (the config is global, independent of the vault):

$ ale toggles          # interactive menu: each toggle with its real approximate cost
ale init --off architect.enrich,digest.map   # disable from the start
ale config set tokens.solutionCache false    # scriptable, one by one

The toggles: Protocol manifest (~6% of the session digest), Architect enrichment in plan_create (~+90 tokens/plan), solution cache (up to ~500 tokens/prompt), per-prompt semantic search (the main saving mechanism — up to ~1500 tokens/prompt, disabling it defeats the point), project map in the digest (up to ~1000 tokens), title index (~200 tokens), and background index rebuild at session close (costs no tokens). All on by default; nothing changes unless you touch them.

Deep project context (ale scan)

ale init --project already runs it; re-run anytime the project changes:

$ ale scan
✓ API: 12 rutas · Env: 8 vars (solo nombres) · Datos: 5 · Convenciones: 4 · .gitignore ✓

Four read-only scanners write compact notes (one line per entry) that the agent receives only when relevant (per-prompt semantic search — the session digest doesn't grow):

• API - — App Router routes with HTTP methods, Pages Router handlers, "use server" actions, Express/Hono routes.
• Env - — environment variable NAMES only, never values (from .env.example / .env / .env.local).
• Datos - — data schema parsed from files (prisma models, drizzle tables, CREATE TABLE in migrations) — no DB connection.
• Convenciones - — package manager, monorepo, TS strict, import aliases, framework.
• Flujo - — what a request goes through: GET /api/x → auth: requirePermiso(...) → datos: getX, createX → audits, plus what the middleware protects. The others say what exists; this one says how it connects.

It never touches your source code, and it guarantees the vault (and personal configs) are gitignored before writing — your notes never reach a shared repo, even if you never ran ale init.

Create a plan from a file

Instead of typing a whole plan into the terminal, keep it as markdown in your repo:

$ ale plan PLAN-migration.md
✓ Plan created: notes/2026-07-11-plan-migrate-to-postgres-17.md
  DoD detected: 3 criteria — backup verified · migrations run clean · app responds on staging

Checkbox lines (- [ ] ...) become the Definition of Done. The agent sees it as an open plan at the start of the next session and resumes it with task_verify.

View the knowledge graph

The graph maintains itself — every vault change (automatic capture, ale add, reindex) regenerates /grafo.html. Double-click it anytime.

ale graph                  # LIVE viewer in your browser: auto-refreshes every 3s
ale graph --out other.html # static copy elsewhere (optional)

Nodes = notes (size = connections, color = type: 🟪 plans, 🟦 verifications, 🟨 lessons, 🟩 sessions…), solid lines = [[wikilinks]], dashed = semantic similarity. Auto-fit, hover highlights neighbors, click shows details, Obsidian-style type filters (prompts hidden by default — they're visual noise). The graph shines from ~20-30 notes on — it's cumulative by design.

In Obsidian: open the vault folder as a vault and use the native graph view — notes carry aliases in their frontmatter so Obsidian resolves the [[wikilinks]]. Old notes are migrated automatically on the next reindex. Note: grafo.html is for your browser; Obsidian doesn't render HTML.

Measure what it saves you

$ ale stats

🧠 Vault: /Users/you/Alexandria (global)
  Notes: 142 {"note":80,"prompt":38,"session":22,"map":2}
  Indexed chunks: 385 · Connections: 91
  Semantic search: active

🏛️  Protocol
  Plans: 6 (1 open) · Verifications: 14 (86% ✓)
  Lessons: 9 (4 reused — each reuse = a debug session that never happened)

📊 Measured (facts, not estimates)
  Context injections: 57
  Solution cache hits: 12 (~9,400 tokens — clearest win: solutions not re-derived)
  Injected tokens (cost): ~21,300

💰 Estimated savings (band, NOT audited)
  ~63,900 – 170,400 tokens
  Assumption: injected context replaces re-reading 3–8× its size in files. Not a
  measurement — the real number depends on what you'd have re-read without the vault.

Audit your vault's health

$ ale audit

🏛️  Alexandria Protocol audit
  Plans: 2 (2 open, 0 completed, 0 failed)
  Verifications: 1 (1 failures) · Lessons: 1
  ⚠ Notes without connections (invisible to vault_related): 1
    - Plan - Migrate to Postgres 17 → connect it with vault_link or [[wikilinks]]

Flags plans without a Definition of Done, failures without lessons, and orphan notes.

Maintenance (you rarely need it)

ale doctor                    # checks & repairs: model, hooks, MCP, index
ale reindex [--force]         # incremental by default
ale consolidate [--days 45]   # ARCHIVES old unused prompts to Alexandria/archive/
                              # — out of the graph and injection, still searchable: nothing is ever lost
ale uninstall                 # removes hooks; your notes stay untouched

Every command self-repairs before running: missing model → downloads it; stale index → updates it.

Supported agents

ale agents                 # list with auto-detection (● = found on your machine)
ale agents all             # register the MCP server in all of them
ale agents cursor,gemini   # just these

| Agent | Config written | Per-project | |---|---|---| | Claude Code | .mcp.json / claude mcp add + hooks | ✓ | | Cursor | ~/.cursor/mcp.json / .cursor/mcp.json | ✓ | | OpenCode | opencode.json | ✓ | | Devin CLI (Cognition) | .devin/mcp_config.local.json | ✓ | | Windsurf (legacy) | ~/.codeium/windsurf/mcp_config.json | — | | Cline | cline_mcp_settings.json (VS Code) | — | | Codex CLI | ~/.codex/config.toml | — | | Gemini CLI | ~/.gemini/settings.json / .gemini/settings.json | ✓ | | VS Code (Copilot) | mcp.json (profile) / .vscode/mcp.json | ✓ | | OpenClaw | ~/.openclaw/openclaw.json | — | | Hermes Agent (Nous) | hermes mcp add / ~/.hermes/config.yaml | — |

Registrations are idempotent merges — your other MCP servers are never touched.

Note: automatic capture via hooks (session digest, per-prompt injection, session save on close) is Claude Code-only — other agents don't have a hook system. There, knowledge flows through the MCP tools, which the agent uses on its own (the protocol instructions travel inside the tool descriptions).

How it works

AI agent session
  │
  ├─ SessionStart ──► injects digest: protocol manifest + project map + open plans   (hooks, Claude Code)
  ├─ Every prompt ──► local hybrid search → injects only what's relevant             (hooks, Claude Code)
  │                   └─ near-identical prompt already solved → injects the solution (cache)
  ├─ MCP tools ─────► vault_search · vault_save · vault_related · vault_link         (all agents)
  │                   plan_create · task_verify · lesson_extract                     (the Protocol)
  └─ Stop/PreCompact► saves the session outcome into the vault                       (hooks, Claude Code)
                                │
                        /Alexandria/*.md   ← markdown, [[wikilinks]]
                        /.vault/                    ← index (regenerable)

• Hybrid search: local embeddings (transformers.js, multilingual e5-small) + BM25, combined with RRF, boosted by recency, usage, and note type (lessons/solutions rank highest).
• Incremental index: only re-processes notes whose mtime changed. The graph is never rebuilt from scratch.
• Deduplication: near-identical prompts don't create new notes — they add hits (and climb the ranking).
• About scores: the e5 model compresses cosines (~0.76–0.86 even for unrelated topics); the ranking is what's reliable, hence the relevance label.

Command reference

| Command | What it does | |---|---| | ale init [--project] [--path ] [--agents ] [--auto\|--manual] [--no-protocol] [--portable] | Installs everything: vault, hooks, MCP, .gitignore, project scan, CLAUDE.md. --portable writes an npx-based MCP command safe to commit for teams | | ale scan | Deep project context: API routes, env var names, data schema (file parse), conventions → compact notes (adaptive cap ≤200 entries, fair sampling in monorepos). Read-only; ensures .gitignore | | ale toggles | Interactive menu to enable/disable every token-spending behavior, each with its real approximate cost. Works before or after install; scriptable via ale init --off | | ale skills:evolve | Generates/updates a living skill for the project from what the vault knows (stack, conventions, app flow, lessons). Always shows a diff and asks before writing | | ale agents [ids] [--project] | List agents / register the MCP server | | ale search [-k n] [--expand] | Hybrid search; --expand pulls graph neighbors. With the Protocol on, appends the chain (plan + verification + lesson) connected to the top hit | | ale add [-c text] [--file ] [-t tags] | Save a note (inline, from file, or stdin) | | ale plan [--title t] | Create a Protocol plan from a .md/.txt file (checkboxes → Definition of Done); surfaces similar past plans + quality hints | | ale graph [--out file.html] [--no-open] | Live local graph viewer | | ale audit | Protocol health: plans without DoD, failures without lessons, orphan notes | | ale stats | Notes, connections, protocol metrics, estimated tokens saved | | ale config [value] | Configuration (e.g. protocol true/false) | | ale skills [-y] [--project] | Recommends & installs Claude skills — semantic (e5) + keyword detection over your vault. Local source configurable with ale config set skills.repo ; otherwise installs from skills.sh | | ale reindex [--force] | Reindex (incremental by default) | | ale consolidate [--days n] [--dry] | Archive old unused prompts — never deletes | | ale doctor [--project] | Check & repair: model, hooks, MCP, index | | ale uninstall [--project] | Remove hooks (notes stay untouched) | | ale --vault | Run any command against another vault |

alexandria works as an alias of ale.

Privacy & security

• **Ev

…

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: UreckChan
• Source: UreckChan/alexandria
• License: MIT
• Homepage: https://alexandria.ureck.dev

Install and usage instructions live in the source repository linked above.