dja

A semantic cache proxy for AI coding tools. dja sits between your coding assistant (e.g., Claude Code) and the Anthropic API, transparently caching responses. When the same (or semantically similar) prompt is sent again, dja returns the cached response instantly, saving time and API costs.

Installation

Quick Install (Linux/macOS)

curl -fsSL https://raw.githubusercontent.com/tomermesser/dja/main/install.sh | sh

Installs to ~/.local/bin and adds it to your PATH automatically.

Cargo

cargo install --git https://github.com/tomermesser/dja

Pre-built Binaries

Download from releases:

• macOS: dja-aarch64-apple-darwin (Apple Silicon)
• Linux: dja-x86_64-unknown-linux-gnu

Verify Installation

dja --version   # Should show current version
dja verify      # Check installation health

Quick Start

dja init          # Downloads embedding model, creates config + database
dja start         # Start the proxy daemon

That's it. dja init sets up shell integration so ANTHROPIC_BASE_URL is automatically configured whenever dja is running. Open a new terminal (or source ~/.zshrc) and your AI tool will route through dja.

How It Works

1. Intercept -- dja proxies all requests to the Anthropic API on 127.0.0.1:9842.
2. Embed -- For eligible requests, dja generates a 384-dimensional embedding of the last user message using a local ONNX model (all-MiniLM-L6-v2). Both single-turn and multi-turn conversations are supported (multi-turn is enabled by default).
3. Lookup -- The embedding is compared against cached entries using cosine similarity via libSQL's vector index. If a match is found above the similarity threshold (default 0.95), the cached response is returned.
4. Store -- On cache miss, the request is forwarded to the upstream API. The response is streamed back to the client and simultaneously buffered for caching.

Cached responses are marked with [cached] at the start of the response text so you can see when a cache hit occurs.

Request coalescing: When identical requests arrive concurrently (e.g., from retries or parallel agents), only one is forwarded upstream. Other waiters are served from cache once the first completes.

Prompt cache optimization: dja automatically injects Anthropic cache_control breakpoints on the system prompt and tool definitions, enabling server-side prompt caching for additional cost savings.

Configuration

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

port = 9842                              # Proxy listen port
upstream = "https://api.anthropic.com"   # Upstream API
threshold = 0.95                         # Cosine similarity threshold
ttl = "30d"                              # Cache entry TTL
max_entries = 10000                      # Maximum cache entries
max_response_size = 1048576              # Max response size to cache (bytes, default 1MB)
log_level = "info"                       # Log level
match_system_prompt = false              # Require system prompt match (false for Claude Code)
multi_turn_caching = true                # Cache multi-turn conversations
auto_cache_control = true                # Auto-inject Anthropic prompt caching breakpoints
request_coalescing = true                # Deduplicate identical in-flight requests

P2P Cache Sharing

dja can share its cache with other machines on the same network. When a prompt misses your local cache, dja checks connected peers before hitting the upstream API.

Setup

P2P is enabled by default when you run dja init. During setup you'll be prompted for a display name, and dja will auto-detect your local IP address.

Both machines need dja installed and running:

# Machine A
curl -fsSL https://raw.githubusercontent.com/tomermesser/dja/main/install.sh | sh
dja init
dja start

# Machine B
curl -fsSL https://raw.githubusercontent.com/tomermesser/dja/main/install.sh | sh
dja init
dja start

Connecting Peers

On Machine A, generate an invite code:

dja p2p invite

This prints a base64 invite code. Copy it and run on Machine B:

dja p2p add <invite-code>

The handshake is mutual — both machines are now connected. If the peer is unreachable at the time, the friend is saved as pending and the handshake completes when you re-run the command.

Managing Peers

dja p2p friends       # List all connected peers
dja p2p status        # Show P2P status (peer ID, address, friend count)
dja p2p remove <id>   # Remove a peer

P2P Configuration

The [p2p] section in ~/.config/dja/config.toml:

[p2p]
enabled = true                  # Enable/disable P2P
peer_id = "dja_6ff7bbdc"        # Auto-generated unique ID
display_name = "MacBook Pro"    # Human-readable name
public_addr = "10.0.1.5:9843"  # Address other peers use to reach you
listen_port = 9843              # P2P API port

If you're using Tailscale or another VPN, update public_addr to your Tailscale IP.

CLI Reference

Command	Description
dja init [--global]	Initialize config, download model, create database
dja start	Start the proxy daemon
dja stop	Stop the proxy daemon
dja status	Show daemon status
dja stats [--json] [--graph]	Show cache statistics or hit graph
dja clear [--older-than 30d]	Clear cache entries
dja config [key] [value]	View or modify configuration
dja test "prompt"	Test embedding and cache lookup
dja export	Export cache as JSON to stdout
dja import <file>	Import cache from JSON file
dja log	Show recent log output
dja verify	Verify installation health
dja monitor	Open live TUI dashboard
dja p2p invite	Generate an invite code for this node
dja p2p add <code>	Add a peer by invite code
dja p2p remove <id>	Remove a peer
dja p2p friends	List all connected peers
dja p2p status	Show P2P status
dja uninstall [--force]	Completely remove dja (binary, data, config, shell hooks)

Uninstall

To completely remove dja from your system:

dja uninstall

This removes the binary, cache database, config, embedding model, and shell integration. Use --force to skip the confirmation prompt.