DCP

Give AI agents permissions. Not your keys.

⬇ Download DCP Desktop — macOS (Apple Silicon & Intel), Windows, Linux. Or start at dcpagent.com.

Use DCP if your AI agent needs to use wallets, API keys, credentials, or user data, but you do not want those secrets sitting in agent configs or .env files.

DCP sits between your agents and your sensitive actions.

Agents ask. You approve, deny, budget, or revoke.

Watch the 15-second demo

Use DCP If

• you run Claude, Cursor, OpenClaw, Hermes, or custom MCP agents
• your agent needs to send, swap, or sign Solana transactions without holding a private key
• your agent needs API keys without reading .env
• you want spending limits for agents
• you want approval before sensitive actions
• you run multiple agents and want one place to manage credentials, permissions, and activity logs

Why DCP?

Agents are useful when they can do real work. Real work needs keys, wallets, credentials, and user data.

The problem is giving an agent raw access is too much trust.

DCP gives agents a permission boundary:

• they ask for what they need
• you approve sensitive actions
• budgets limit damage
• private keys stay in your vault
• every action is logged

What It Does

• stores wallets, API keys, and user data locally
• lets agents request access through MCP
• asks you before sensitive actions
• signs transactions without exposing private keys
• enforces per-agent budgets
• logs what agents did

DCP exposes vault permissions through MCP, so Claude Desktop, Cursor, OpenClaw, Hermes, and custom agents can request approved actions without reading raw secrets directly.

5-Minute Quickstart

By the end, your agent can ask DCP for your Solana wallet address.

Local Agent

1. Download DCP Desktop from dcpagent.com.
2. Create and unlock your vault.
3. Create a Solana wallet.
4. Open Connect and add Claude Desktop, Cursor, Hermes, or another MCP agent.
5. Restart your agent app.

Then ask your agent:

What is my Solana wallet address from DCP?

Remote Agent

For OpenClaw, Hermes, or any agent running on a VPS, create a remote invite in DCP Desktop and run the generated command on the VPS:

curl -fsSL https://dcpagent.com/install.sh | sudo bash -s -- 'dcp_vps_v1_...'

Approve the verification phrase in Desktop. The installer pairs the VPS, starts DCP as a systemd service, and configures OpenClaw or Hermes when either is detected.

Good install output ends with:

DCP service health: ok
Hermes config written: yes
Hermes config verified: yes

For OpenClaw, start a fresh chat/session. For Hermes, run:

/reload-mcp

Cloud Agent (paste a link)

For agents you do not host — Claude.ai, ChatGPT, or hosted OpenClaw/Hermes — there is nothing to install. In DCP Desktop, generate a connect link and paste it into the agent:

dcp_connect_v1_...

The link is single-use, expires in ~10 minutes, and carries no permissions. It pins your vault's public key so the relay cannot impersonate your vault. The agent reaches your on-device vault through the DCP relay (an end-to-end-encrypted MCP facade), and you approve the connection on your device before anything is granted. Scopes, budgets, and revocation stay in your vault.

CLI

Use the CLI to create and manage vault data:

npm install -g @dcprotocol/vault @dcprotocol/agent
dcp init
dcp create-wallet --chain solana
dcp add credentials.api.openai
dcp list

Once a local stdio MCP client is configured, it runs:

{
  "command": "dcp-agent",
  "args": ["run", "--mode", "mcp", "--agent", "claude_desktop"]
}

Try These Prompts

What is my Solana wallet address from DCP?

Read my OpenAI credential from DCP.

Check if sending 0.01 SOL is within my DCP budget.

Request approval to sign a Solana transaction.

How It Works

Claude / Cursor / OpenClaw / Hermes
        |
        v
    dcp-agent
        |
        v
 Local DCP vault
        |
        v
 approve / deny / budget / revoke
        |
        v
 wallets, API keys, identity data

The agent asks for an action. The vault checks policy. If approval is needed, DCP creates a consent request. The agent gets only the result, not the raw private key.

What Approval Shows

When an agent requests a sensitive action, DCP shows:

• which agent is asking
• what action it wants
• amount, chain, and destination for payments
• whether the request is within budget
• approve or deny controls

Screenshots

Real-time approval

MCP connection

Multiple agents

Security Model

DCP is designed around least privilege.

• private keys never leave the local vault
• agents receive results, not raw private keys
• sensitive actions can require explicit approval
• budgets limit automated spending
• scopes control which data an agent can request
• access can be revoked per agent
• sensitive activity is logged

Supported Today

• Solana transfers and swaps (build, budget-check, approve, sign, submit)
• Solana transaction and message signing
• wallet balances, token search, and transaction history
• scoped vault reads and writes
• API credential storage
• budget checks
• stdio MCP for Claude Desktop, Cursor, Hermes, and similar clients
• HTTP MCP for local or custom agents
• Desktop approvals
• Telegram approvals
• remote/VPS OpenClaw and Hermes agents through relay

MCP Tools Exposed

DCP exposes these MCP tools:

• vault_get_address
• vault_budget_check
• vault_scope_guide
• vault_sign_tx
• vault_sign_message
• vault_sign_x402
• vault_read
• vault_write

What Agents Can Request

• wallet address
• transaction signature
• message signature
• API credential access
• identity or profile data
• budget check

What Agents Cannot Do

• read private keys
• bypass approval
• access scopes they were not granted
• spend past configured limits
• silently export the vault

Desktop App

Prefer a GUI? Download DCP Desktop from dcpagent.com.

Desktop Flow

DCP Desktop is the easiest way to get started.

1. Download DCP Desktop for macOS, Windows, or Linux.
2. Create a vault with a password/passphrase.
3. Save your recovery phrase safely.
4. DCP creates a Solana wallet for you.
5. Add private data or credentials in the Data tab.
6. Connect local agents like Claude, Cursor, VS Code, OpenClaw, Hermes, or any MCP client.
7. Set permissions per agent.
8. Approve, deny, budget, revoke, and audit every action agents ask for.

Remote Agents

For a VPS or remote agent, create an invite in DCP Desktop and run the generated command:

curl -fsSL https://dcpagent.com/install.sh | sudo bash -s -- 'dcp_vps_v1_...'

That one command:

• installs a private DCP Node runtime when the VPS does not already have compatible Node.js
• installs @dcprotocol/agent under /var/lib/dcp-agent
• creates and starts dcp-agent.service
• pairs the VPS with your Desktop vault through the relay
• starts HTTP MCP on the VPS
• configures OpenClaw and Hermes when either is detected

Host-native Hermes uses ~/.hermes/config.yaml. Docker Hermes uses /opt/data/config.yaml inside the Hermes container and gets a Docker-reachable DCP URL instead of 127.0.0.1.

If Hermes is not detected or automatic config is not verified, run these on the remote host as the Hermes user:

hermes config set mcp_servers.dcp.url http://127.0.0.1:8420/mcp
hermes config set mcp_servers.dcp.enabled true
hermes config set mcp_servers.dcp.tools.prompts false
hermes config set mcp_servers.dcp.tools.resources false

Then run /reload-mcp in Hermes or restart Hermes.

If OpenClaw and Hermes run on the same VPS, they can share the same DCP MCP endpoint. The installer preserves the OpenClaw config shape and only changes the DCP service runtime so systemd runs the installed package directly.

Packages

Package	Purpose
@dcprotocol/vault	Vault CLI and local vault server
@dcprotocol/agent	MCP, HTTP MCP, and remote sidecar runtime
@dcprotocol/core	Crypto, storage, wallet, policy, and shared types
@dcprotocol/relay	Encrypted relay for remote agents
@dcprotocol/telegram	Telegram approval service
@dcprotocol/desktop	Desktop vault app

Developer Checks

pnpm install
pnpm -r run typecheck
pnpm -r run test
pnpm -r run build
node scripts/publish-guard.mjs
./scripts/test-security.sh