epicsagas/alcove

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

→ Quick start

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.

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.

Why Alcove

CLAUDE.md / AGENTS.md is for agent behavior — recurring corrections, coding conventions, session instructions. Keep it under 200 lines.

Alcove is for project knowledge — architecture, decisions, runbooks, PRDs. Agents search for what they need; nothing is pre-loaded into context.

CLAUDE.md | AGENTS.md            ← agent rules, coding conventions, recurring corrections
~/.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

Quick start

# macOS
brew install epicsagas/tap/alcove

# Linux / Windows — pre-built binary (fast, no compilation)
# Requires Rust: https://rustup.rs
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

Then run setup:

alcove setup

# Verify everything is working
alcove doctor

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

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.

Troubleshooting

Agent can't find Alcove tools Run alcove setup again — it re-registers the MCP server for all configured agents. Then start a new agent session (registration takes effect on next session start).

Search returns no results The index may not be built yet. Run alcove index to build it, then try again.

403 Unauthorized from background server ALCOVE_TOKEN is not set in your shell. Run alcove token to print it, then add export ALCOVE_TOKEN="..." to your shell profile and reload.

alcove doctor reports issues Follow the suggestions printed by doctor — it checks binary location, MCP registration, index state, and optional dependencies like pdftotext.

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.

MCP Tools

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:

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

Running a persistent background server eliminates cold-start latency (2–5s ONNX model load) on every new agent session. alcove setup enables this by default (macOS login item).

# Enable and start (persists across reboots — macOS)
alcove enable

# Lifecycle
alcove stop / start / restart

# Disable and remove login item
alcove disable

The server uses a bearer token for authentication — auto-generated during alcove setup and stored in config.toml. Your existing MCP config (command: alcove) stays unchanged; the stdio process auto-detects the running server and proxies to it.

# Check or share the token
alcove token

# Set in shell profile (setup does this automatically)
export ALCOVE_TOKEN="alcove-..."

Token priority: --token flag > ALCOVE_TOKEN env var > config.toml.

Logs are written to ~/.alcove/logs/. On startup, run alcove doctor to verify the server is reachable.

Proxy mode (background server running):
  Agent ──stdio──→ alcove (thin proxy, <10 MB, instant)
                     └──HTTP──→ alcove serve (always warm)
                                 ├─ BM25 index (loaded)
                                 ├─ ONNX embedding model (loaded)
                                 ├─ HNSW vector index (loaded)
                                 └─ hybrid search ready (~5ms)

Direct mode (no background server):
  Agent ──stdio──→ alcove (full engine, one process per session)
                     ├─ load ONNX embedding model (2-5s cold start)
                     ├─ open BM25 index
                     └─ hybrid search ready (after warm-up)

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

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:

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:

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 = 57384                  # 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

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.

# 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.

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