epicsagas/alcove

Your AI agent doesn't know your project. Alcove fixes that.

English · 한국어 · 日本語 · 简体中文 · Español · हिन्दी · Português · Deutsch · Français · Русский

Alcove is an MCP server that gives AI coding agents on-demand access to your private project docs — without dumping everything into the context window, without leaking docs into public repos, and without per-project config for every agent you use.

Demo

Claude, Gemini, Codex — search · switch projects · global search · validate & generate. One setup.

CLI demo

alcove search · project switch · --scope global · alcove validate

The problem

You have two bad options.

Option A: Put docs in CLAUDE.md / AGENTS.md Every file gets injected into the context window on every run. Works for short conventions. Breaks down with real project docs. 10 architecture files = context bloat = slower, more expensive, less accurate responses.

Option B: Don't put docs in Your agent invents requirements you already documented. It ignores constraints from decisions you already made. It asks you to explain the same things every session.

Neither option scales. Now multiply it across 5 projects and 3 agents, each configured differently. Every time you switch, you lose context.

How Alcove solves this

Alcove doesn't inject your docs. Agents search for what they need, when they need it.

~/projects/my-app $ claude "how is auth implemented?"

  → Alcove detects project: my-app
  → BM25 search: "auth" → ARCHITECTURE.md (score: 0.94), DECISIONS.md (score: 0.71)
  → Agent gets the 2 most relevant docs, not all 12

~/projects/my-api $ codex "review the API design"

  → Alcove detects project: my-api
  → Same doc structure, same access pattern
  → Different project, zero reconfiguration

Switch agents anytime. Switch projects anytime. The document layer stays standardized.

The right split

CLAUDE.md / AGENTS.md is for agent behavior: repeated mistakes to avoid, coding conventions, and session-specific instructions. Keep it under 200 lines.

Alcove is for project knowledge: architecture, decisions, runbooks, API docs, and anything else your agent needs to understand — but not necessarily on every run.

The pattern:

CLAUDE.md | AGENTS.md                             ← agent rules, coding conventions, recurring corrections
~/.config/alcove/docs/my-app/
  ARCHITECTURE.md                      ← tech stack, data model, system design
  DECISIONS.md                         ← why X was chosen over Y
  DEBT.md                              ← known issues, workarounds
  ...                                  ← agent searches here when it needs context

Agents call search_project_docs("auth flow") and get the 2 most relevant docs — not all 12. Nothing hits the context window unless it's actually needed.

Why Alcove

Why not just use CLAUDE.md? Short conventions and agent behaviors belong there. Project documentation — architecture, decisions, runbooks, PRDs — doesn't scale in a context file. Alcove is not a replacement; it's the layer CLAUDE.md was never meant to be.

Without Alcove	With Alcove
Docs in CLAUDE.md bloat context on every run	BM25 search — agents pull only what they need
Internal docs scattered across Notion, Google Docs, local files	One doc-repo, structured by project
Each AI agent configured separately for doc access	One setup, all agents share the same access
Switching projects means re-explaining context	CWD auto-detection, instant project switch
Agent search returns random matching lines	Ranked results — best matches first, one result per file
"Search all my notes about OAuth" — impossible	Global search across every project in one query
Sensitive docs sitting in project repos	Private docs on your machine, never in public repos
Doc structure differs per project and team member	policy.toml enforces standards across all projects
No way to check if docs are complete	validate catches missing files, empty templates, missing sections
Stale docs with broken links or WIP markers go unnoticed	lint detects broken links, orphans, and stale markers automatically
Notes from Obsidian or other tools stay siloed	promote brings any note into your doc-repo with one command

Quick start

# macOS
brew install epicsagas/tap/alcove

# Linux / Windows — pre-built binary (fast, no compilation)
cargo install cargo-binstall
cargo binstall alcove

# Any platform — build from source
cargo install alcove

# Clone and build
git clone https://github.com/epicsagas/alcove.git
cd alcove
make install

alcove setup

Claude Code Plugin

If you use Claude Code, you can install Alcove as a plugin — it auto-installs the binary and registers the MCP server in one step:

claude plugin install epicsagas/alcove

This runs a SessionStart hook that:

1. Installs the alcove binary if not found (via brew / cargo-binstall / cargo)
2. Runs alcove setup to register the MCP server

Optional dependencies

Tool	Purpose	Install
pdftotext (poppler)	Full PDF text extraction — required for PDF search	macOS: brew install poppler · Debian/Ubuntu: apt install poppler-utils · Fedora: dnf install poppler-utils · Windows: poppler for Windows

Without pdftotext, Alcove falls back to a built-in PDF parser which may fail on some files. Run alcove doctor to check your setup.

Note: Pre-built binaries are available for Linux (x86_64), macOS (Apple Silicon & Intel), and Windows.

setup walks you through everything interactively:

1. Where your docs live
2. Which document categories to track
3. Preferred diagram format
4. Embedding model for hybrid search
5. HTTP server — host, port, auto-generated bearer token, and login item registration
6. Which AI agents to configure (MCP + skill files)

Re-run alcove setup anytime to change settings. It remembers your previous choices.

---

How it works

flowchart LR
    subgraph Projects["Your projects"]
        A1["my-app/\n  src/ ..."]
        A2["my-api/\n  src/ ..."]
    end

    subgraph Docs["Your private docs (one repo)"]
        D1["my-app/\n  PRD.md\n  ARCH.md"]
        D2["my-api/\n  PRD.md\n  ..."]
        P1["policy.toml"]
    end

    subgraph Agents["Any MCP agent"]
        AG["Claude Code · Cursor\nGemini CLI · Codex · Copilot\n+4 more"]
    end

    subgraph MCP["Alcove MCP server"]
        T["search · get_file\noverview · audit\ninit · validate"]
    end

    A1 -- "CWD detected" --> D1
    A2 -- "CWD detected" --> D2
    Agents -- "stdio (proxy → HTTP)" --> MCP
    MCP -- "scoped access" --> Docs

Your docs are organized in a separate directory (DOCS_ROOT), one folder per project. Alcove manages docs there and serves them to any MCP-compatible AI agent over stdio. When the background HTTP server is running (via alcove enable), the stdio process acts as a thin proxy — forwarding requests to the warm server for instant response with zero cold-start. Without the background server, it loads the full engine on each session.

What it does

• On-demand doc retrieval — agents search and retrieve; nothing is pre-loaded into context
• BM25 ranked search — fast full-text search powered by tantivy; most relevant docs first, auto-indexed, falls back to grep
• One doc-repo, multiple projects — private docs organized by project, managed in a single place
• One setup, any agent — configure once, every MCP-compatible agent gets the same access
• Auto-detects your project from CWD — no per-project config needed
• Scoped access — each project only sees its own docs
• Cross-project search — search across all projects at once with scope: "global"
• Private docs stay private — docs never touch your public repo, runs entirely on your machine
• Persistent HTTP server — optional background server eliminates cold-start latency; agents connect via HTTP for instant response
• Standardized doc structure — policy.toml enforces consistent docs across all projects and teams
• Cross-repo audit — finds internal docs misplaced in your project repo, suggests fixes
• Document validation — checks for missing files, unfilled templates, required sections
• Semantic lint — detects broken wikilinks, orphan files, stale WIP/DRAFT markers, and date claims that are 2+ years old
• External vault promotion — bring a note from Obsidian (or any vault) into your alcove doc-repo with one command; auto-routes to the right project
• Knowledge base vaults — create, link, and search independent knowledge bases (separate from project docs); link Obsidian vaults directly
• Works with 9+ agents — Claude Code, Cursor, Claude Desktop, Cline, OpenCode, Codex, Copilot, Antigravity, Gemini CLI

MCP Tools

Tool	What it does
get_project_docs_overview	List all docs with classification and sizes
search_project_docs	Smart search — auto-selects BM25 ranked or grep, supports scope: "global" for cross-project search
get_doc_file	Read a specific doc by path (supports offset/limit for large files)
list_projects	Show all projects in your docs repo
audit_project	Cross-repo audit — scans doc-repo and local project repo, suggests actions
init_project	Scaffold docs for a new project (internal + external docs, selective file creation)
validate_docs	Validate docs against team policy (policy.toml)
rebuild_index	Rebuild the full-text search index (usually automatic)
check_doc_changes	Detect added, modified, or deleted docs since last index build
lint_project	Semantic lint — broken links, orphan files, stale markers, stale date claims
promote_document	Copy or move a file from an external vault into the alcove doc-repo
search_vault	Search knowledge base vaults — separate from project docs, for research and reference
list_vaults	List all knowledge base vaults with document counts

CLI

alcove              Start MCP server (agents call this)
alcove setup        Interactive setup — re-run anytime to reconfigure
alcove doctor       Check the health of your alcove installation
alcove validate     Validate docs against policy (--format json, --exit-code)
alcove lint         Semantic lint — broken links, orphans, stale markers (--format json)
alcove promote      Bring a file from an external vault into your doc-repo
alcove index        Update the search index (incremental — only changed files)
alcove rebuild      Rebuild the search index from scratch (use after schema changes)
alcove search       Search docs from the terminal
alcove serve        Start HTTP RAG server for external access
alcove enable       Register as macOS login item and start background server
alcove disable      Unregister from login items and stop server
alcove start        Start the background server
alcove stop         Stop the background server
alcove restart      Restart the background server
alcove token        Print the bearer token for team sharing
alcove uninstall    Remove skills, config, and legacy files

alcove vault create   Create a new knowledge base vault
alcove vault link     Link an external directory as a vault (e.g., Obsidian)
alcove vault list     List all vaults with document counts
alcove vault remove   Remove a vault (symlinks: remove link only)
alcove vault add      Add a document to a vault
alcove vault index    Build search index for vaults
alcove vault rebuild  Rebuild vault search index from scratch

Lint

# Lint the current project (auto-detected from CWD)
alcove lint

# Lint a specific project by name
alcove lint --project my-app

# Machine-readable output for CI
alcove lint --format json

Lint checks four things:

Check	What it catches
broken-link	[[wikilinks]] and [text](path) pointing to missing files
orphan	Files that no other document links to
stale-marker	WIP / TODO / FIXME / DRAFT / DEPRECATED markers
stale-date	Year mentions that are 2+ years old (e.g. "as of 2022")

Promote

# Copy a note from Obsidian into your doc-repo (auto-routes to matching project)
alcove promote ~/my-brain/Projects/auth-notes.md

# Route to a specific project
alcove promote ~/my-brain/Projects/auth-notes.md --project my-app

# Move instead of copy
alcove promote ~/my-brain/Projects/auth-notes.md --mv

Files with no matching project land in inbox/ for manual review.

Background Server

Alcove can run as a persistent HTTP RAG server, accessible via REST API. This is useful for external integrations, dashboards, or non-MCP clients. When enabled, the MCP stdio process automatically proxies to the warm server — eliminating cold-start latency (ONNX model load, index open) on every new session.

# Start the server in the foreground
alcove serve                       # default: 127.0.0.1:8080
alcove serve --port 9090           # custom port
alcove serve --host 0.0.0.0       # listen on all interfaces

The server uses a bearer token for authentication. During alcove setup, a token is auto-generated and stored in config.toml. You can also pass one explicitly with --token or the ALCOVE_TOKEN environment variable.

Token management

# Print the stored token (for sharing with teammates)
alcove token

# Teammates set it in their shell profile:
export ALCOVE_TOKEN="alcove-a3f7b2e14d5c..."

Tokens are resolved with priority: --token flag > ALCOVE_TOKEN env var > config.toml.

macOS Login Item (launchd)

Register Alcove as a macOS login item so the HTTP server starts automatically on login and stays running in the background. This is the default during alcove setup — the setup wizard asks whether to enable it (default: Yes).

# Register and start (persists across reboots)
alcove enable

# Lifecycle management
alcove stop         # stop the server
alcove start        # start it again
alcove restart      # stop + start

# Unregister (stops server and removes login item)
alcove disable

This installs a LaunchAgent at ~/Library/LaunchAgents/com.epicsagas.alcove.plist. Logs are written to ~/.alcove/logs/.

Hybrid Proxy Mode

Agents always connect to alcove via stdio (the MCP standard). When the background HTTP server is running, the stdio process acts as a thin proxy — it forwards JSON-RPC messages to the warm server over HTTP instead of loading the search engine itself. This eliminates cold-start latency (ONNX model load, index open) on every new agent session.

With background server (proxy mode):
  Agent ──stdio──→ alcove (thin proxy)
                     │ stdin → HTTP POST /mcp
                     │ HTTP response → stdout
                     └──HTTP──→ alcove serve (always warm)
                                 ├─ BM25 index (loaded)
                                 ├─ ONNX embedding model (loaded)
                                 ├─ HNSW vector index (loaded)
                                 └─ hybrid search ready (~5ms)

Without background server (direct mode):
  Agent ──stdio──→ alcove (full engine)
                     ├─ load ONNX embedding model (2-5s cold start)
                     ├─ open BM25 index
                     ├─ build HNSW vector index
                     └─ hybrid search ready (after warm-up)

On startup, the stdio process checks GET /health on the configured host/port. If the server responds, it enters proxy mode automatically — no configuration change needed. The JSON-RPC payload is identical in both modes; only the transport changes internally.

Search

Alcove automatically picks the best search strategy. When the search index exists, it uses BM25 ranked search (powered by tantivy) for relevance-scored results. When it doesn't, it falls back to grep. You never have to think about it.

Hybrid Search (RAG)

Alcove supports Hybrid Search which combines BM25 with Vector Similarity Search (powered by fastembed).

During alcove setup, you can choose an embedding model and download it immediately. You can also manage models manually:

# Set and download an embedding model
alcove model set MultilingualE5Small
alcove model download

# Check model status
alcove model status

Choosing a model

Model	Disk	Dim	Languages	Best for
SnowflakeArcticEmbedXSQ	15 MB	384	English	CI, resource-constrained environments
SnowflakeArcticEmbedXS	30 MB	384	English	Fast English-only indexing
SnowflakeArcticEmbedSQ	65 MB	384	English	Balanced quality + size (English)
SnowflakeArcticEmbedS	130 MB	384	English	Good English recall
MultilingualE5Small	235 MB	384	100+ languages	Default — multilingual / mixed-language projects
SnowflakeArcticEmbedMQ	200 MB	768	English	High quality, quantized
SnowflakeArcticEmbedM	400 MB	768	English	Best English recall
MultilingualE5Base	555 MB	768	100+ languages	Better multilingual quality
MultilingualE5Large	2.2 GB	1024	100+ languages	Maximum multilingual quality
BGEM3	2.3 GB	1024	100+ languages	State-of-the-art multilingual

Q (quantized) variants use int8 quantization — ~50% smaller on disk, slightly lower recall, no meaningful accuracy loss for typical document search. Use the XSQ/SQ/MQ variants when memory is a constraint.

Once a model is downloaded and ready, Alcove will automatically use Hybrid Search for both CLI search and agent-based MCP tools. This is particularly effective for multilingual projects and complex semantic queries.

# Search the current project (auto-detected from CWD)
alcove search "authentication flow"

# Force grep mode if you want exact substring matching
alcove search "FR-023" --mode grep

The index builds automatically in the background when the MCP server starts, and rebuilds when it detects file changes. No cron jobs, no manual steps.

How it works for agents: agents just call search_project_docs with a query. Alcove handles the rest — ranking, deduplication (one result per file), cross-project search, and fallback. The agent never needs to choose a search mode.

Index lifecycle

Understanding when to run alcove index vs alcove rebuild:

Command	What it does	When to use
alcove index	Incremental update — only processes new/changed files	Default: run after adding or editing docs
alcove rebuild	Full rebuild — drops and recreates all index data	After changing embedding models, or after index corruption

First-time setup:

# Step 1: BM25 search is ready immediately after setup
alcove index            # builds full-text index (no model needed)

# Step 2: Enable Hybrid Search (optional but recommended)
alcove model set MultilingualE5Small
alcove model download   # ~235 MB download

# Step 3: Build vector index for all existing docs
alcove rebuild          # one-time full rebuild with embeddings
                        # ⚠ peak RAM = model size + corpus vectors (see note below)

# After this: incremental updates just work
alcove index            # fast — only re-embeds changed files

Switching models:

alcove model set SnowflakeArcticEmbedS   # change model
alcove rebuild                            # required: vectors are model-specific

Memory during rebuild:
Peak RAM = model size + all document vectors held in RAM while building the HNSW graph. For MultilingualE5Small with ~3,500 docs, expect ~700 MB peak. This is structural — after rebuild completes, steady-state drops to ~50–200 MB depending on your [memory] config. You can reduce steady-state further with lower max_hnsw_cache and shorter model_unload_secs.

Global search

Every architecture decision, every runbook, every project note — searchable across all your projects at once.

# Search across ALL projects
alcove search "rate limiting patterns" --scope global
alcove search "OAuth token refresh" --scope global

Agents can do the same with scope: "global" in search_project_docs. One query, every project.

Project detection

By default, Alcove detects the current project from your terminal's working directory (CWD). You can override this with the MCP_PROJECT_NAME environment variable:

MCP_PROJECT_NAME=my-api alcove

This is useful when your CWD doesn't match a project name in your docs repo.

Document policy

Define team-wide documentation standards with policy.toml in your docs repo:

[policy]
enforce = "strict"    # strict | warn

[[policy.required]]
name = "PRD.md"
aliases = ["prd.md", "product-requirements.md"]

[[policy.required]]
name = "ARCHITECTURE.md"

  [[policy.required.sections]]
  heading = "## Overview"
  required = true

  [[policy.required.sections]]
  heading = "## Components"
  required = true
  min_items = 2

Policy files are resolved with priority: project (<project>/.alcove/policy.toml) > team (DOCS_ROOT/.alcove/policy.toml) > built-in default (from your config.toml core files). This ensures consistent doc quality across all your projects while allowing per-project overrides.

Document classification

Alcove classifies docs into tiers:

Classification	Where it lives	Examples
doc-repo-required	Alcove (private)	PRD, Architecture, Decisions, Conventions
doc-repo-supplementary	Alcove (private)	Deployment, Onboarding, Testing, Runbook
reference	Alcove reports/ folder	Audit reports, benchmarks, analysis
project-repo	Your GitHub repo (public)	README, CHANGELOG, CONTRIBUTING

The audit tool scans both your doc-repo and local project directory, then suggests actions — like generating a public README from your private PRD, or pulling misplaced reports back into Alcove.

Configuration

Config lives at ~/.config/alcove/config.toml:

docs_root = "/Users/you/documents"

[core]
files = ["PRD.md", "ARCHITECTURE.md", "PROGRESS.md", "DECISIONS.md", "CONVENTIONS.md", "SECRETS_MAP.md", "DEBT.md"]

[team]
files = ["ENV_SETUP.md", "ONBOARDING.md", "DEPLOYMENT.md", "TESTING.md", ...]

[public]
files = ["README.md", "CHANGELOG.md", "CONTRIBUTING.md", "SECURITY.md", ...]

[diagram]
format = "mermaid"

[server]
host = "127.0.0.1"          # bind address (0.0.0.0 for all interfaces)
port = 8080                  # listen port
token = "alcove-a3f7b2..."   # auto-generated bearer token

[memory]
reader_ttl_secs   = 300   # evict idle IndexReader after N seconds (0 = never)
max_cached_readers = 1    # max concurrent IndexReader instances in RAM
model_unload_secs  = 600  # unload embedding model after N seconds of inactivity (0 = never)
max_hnsw_cache     = 3    # max HNSW graphs held in memory simultaneously

All of this is set interactively via alcove setup. You can also edit the file directly.

Memory usage note: During initial indexing or a full rebuild, Alcove loads the embedding model (~235–500 MB) and holds all document vectors in RAM while constructing the HNSW graph — peak usage scales with corpus size and is unavoidable for that operation. The [memory] settings above control steady-state RAM after indexing is complete.

File lists are fully customizable — add any filename to any category, or move files between categories to match your team's workflow:

[core]
files = ["PRD.md", "ARCHITECTURE.md", "DECISIONS.md", "MY_SPEC.md"]  # added custom doc

[public]
files = ["README.md", "CHANGELOG.md", "PRD.md"]  # PRD exposed as public for this project

Supported agents

Agent	MCP	Skill
Claude Code	~/.claude.json	~/.claude/skills/alcove/
Cursor	~/.cursor/mcp.json	~/.cursor/skills/alcove/
Claude Desktop	platform config	—
Cline (VS Code)	VS Code globalStorage	~/.cline/skills/alcove/
OpenCode	~/.config/opencode/opencode.json	~/.opencode/skills/alcove/
Codex CLI	~/.codex/config.toml	~/.codex/skills/alcove/
Copilot CLI	~/.copilot/mcp-config.json	~/.copilot/skills/alcove/
Antigravity	~/.gemini/antigravity/mcp_config.json	—
Gemini CLI	~/.gemini/settings.json	~/.gemini/skills/alcove/

Agents with skill support activate Alcove automatically when you ask about project architecture, conventions, decisions, or status. They can also be invoked explicitly:

/alcove                          Summarize current project docs and status
/alcove search auth flow         Search docs for a specific topic
/alcove what conventions apply?  Ask a doc question directly

Supported languages

The CLI automatically detects your system locale. You can also override it with the ALCOVE_LANG environment variable.

Language	Code
English	en
한국어	ko
简体中文	zh-CN
日本語	ja
Español	es
हिन्दी	hi
Português (Brasil)	pt-BR
Deutsch	de
Français	fr
Русский	ru

# Override language
ALCOVE_LANG=ko alcove setup

Update

# Homebrew
brew upgrade epicsagas/tap/alcove

# cargo-binstall
cargo binstall alcove

# From source
cargo install alcove

Uninstall

alcove uninstall          # remove skills & config
cargo uninstall alcove    # remove binary

Knowledge Base Vaults

Beyond project documentation, Alcove supports independent knowledge base vaults for research notes, reference materials, and curated knowledge that LLMs can search.

# Create a vault for AI research notes
alcove vault create ai-research

# Link an existing Obsidian vault (no copying — indexes in place)
alcove vault link my-obsidian ~/Obsidian/research

# Add a document
alcove vault add ai-research ~/Downloads/transformer-survey.md

# Build the vault search index
alcove vault index

# Search from CLI
alcove search "attention mechanism" --vault ai-research

# Agents search via MCP
search_vault(query="attention mechanism", vault="ai-research")

# Search ALL vaults at once
search_vault(query="transformer", vault="*")

Vaults are completely isolated from project docs — separate indexes, separate caches, separate search. Your coding agent's project doc search is never affected by vault activity.

Feature	Project docs	Vaults
Purpose	Per-project documentation	General knowledge base
Storage	~/.alcove/docs/	~/.alcove/vaults/
Index	Shared project index	Independent per-vault index
Cache	PROJECT_READER_CACHE	VAULT_READER_CACHE
Search	search_project_docs	search_vault
Symlink	No	Yes (link external dirs)

Ecosystem

obsidian-forge

Alcove pairs naturally with obsidian-forge, an Obsidian vault generator and automation daemon. Use obsidian-forge to build and strengthen your knowledge graph in Obsidian, then promote notes into alcove with alcove promote — your AI agents get ranked, scoped search over your project knowledge base without any context bloat.

obsidian-forge (personal knowledge)   →   alcove promote   →   alcove (project docs)
  vault / inbox / graph                    one command           BM25 + vector search

Contributing

Bug reports, feature requests, and pull requests are welcome. Please open an issue on GitHub to start a discussion.

License

Apache-2.0