llms.txt
# DCP: Local control app for AI agents
## What is DCP?
DCP is a local-first control app for AI agents.
It lets AI agents use wallets, credentials, identity, and private data inside rules set by the user. Agents can request actions like signing a Solana transaction, paying for an API, using a credential, or reading approved data. DCP checks the rules locally. Allowed actions can run automatically. Unusual or over-limit actions ask for approval through Desktop, Telegram, or the DCP mobile app.
DCP runs on the user's machine. Wallet private keys stay encrypted locally and never enter the agent runtime.
Short version: DCP lets AI agents use sensitive things without giving them full control.
## One-line description
DCP is a local-first, non-custodial control app and MCP permission server for AI agents.
## Main positioning
One App. Any agent. You stay in control.
Agents get access. Users keep control.
## What DCP controls
DCP controls how AI agents use:
- Solana wallets
- API keys and credentials
- Identity data
- Private or project data
- x402 / USDC payments
- Remote-agent access
- Per-agent budgets, limits, approvals, and logs
## When does DCP matter?
DCP matters when an AI agent needs real authority.
Use DCP when an agent needs to:
- sign a Solana transaction
- pay for an API, dataset, inference, service, or another agent
- use a paid or production credential such as OpenAI, Anthropic, Stripe, AWS, GitHub, or custom API keys
- read private or identity data
- run on a VPS or cloud machine while keys stay local
- take an action that should be limited, logged, or approved
If an agent only chats or edits local files, it may not need DCP. If an agent touches wallets, credentials, identity, payments, or private data, DCP is useful.
## Who is DCP for?
DCP is for people and teams running AI agents that need controlled access to sensitive resources.
Main users:
- Solana and x402 agent builders
- OpenClaw, Hermes, and VPS agent operators
- AI agent framework builders using MCP, OpenAI Agents SDK, LangGraph, CrewAI, AutoGen, LangChain, or custom Node/Python agents
- No-code automation power users using n8n, Zapier, Make, Gumloop, or similar tools
- Multi-agent users running Claude, Cursor, OpenClaw, Hermes, and custom agents with different access rules
## How DCP works
1. The user adds wallets, credentials, identity, or private data to DCP.
2. The user connects agents through MCP, Desktop integration, or remote agent setup.
3. The user sets rules for each agent.
4. The agent requests an action.
5. DCP checks the request locally against scope, limits, budget, session, and approval rules.
6. If allowed, the action runs automatically.
7. If unusual or over the limit, DCP asks for approval through Desktop, Telegram, or the DCP mobile app.
8. The agent receives the approved result, not full control.
## Example
An OpenClaw agent running on a VPS needs to pay for API access.
Instead of storing a private key or API key on the VPS, the agent calls DCP. DCP checks the user's rules. If the payment is under the allowed limit, DCP approves it automatically. If it crosses the limit, the user receives an approval request. The key stays local.
## Main product idea
DCP solves the locked-out vs overpowered agent problem.
Without DCP, users often choose between:
- locking the agent out, making it safe but useless
- giving the agent raw secrets, making it useful but risky
- approving every action manually, making the agent non-autonomous
DCP gives a fourth option: controlled access.
## Core features
- Local-first vault
- Solana wallet signing
- x402 payment support
- Per-agent rules
- Per-agent budgets and limits
- Desktop approvals
- Telegram approvals through @dcpagentBot
- Remote/VPS agent support
- MCP permission server
- Scoped access to credentials and private data
- Audit logs
- Instant revocation
- Apache-2.0 open source license
## Security model
DCP is designed so the user keeps control.
- Wallet private keys stay encrypted on the user's machine.
- Agents do not receive raw wallet private keys.
- Agents operate inside per-agent rules.
- Allowed actions can run automatically.
- Over-limit or unusual actions require approval.
- Every action is logged.
- Access can be revoked anytime.
- DCP is open source and local-first.
Encryption: XChaCha20.
Master key storage: OS keychain.
## What can DCP store?
DCP can store or manage:
- Solana wallets
- API keys and service credentials
- OpenAI keys
- Anthropic keys
- Stripe keys
- AWS credentials
- GitHub tokens
- identity data such as name, email, phone, address, and profile details
- private notes
- project data
- custom fields
Important distinction:
- For wallet signing, agents never receive the private key.
- For API keys and credentials, DCP gives agents scoped, approved, logged, and revocable access.
## How is DCP different from agent wallets?
Agent-wallet products usually create a new wallet or account for the agent.
DCP lets agents request controlled use of the wallets and credentials the user already trusts.
Agent wallets give agents accounts.
DCP gives agents controlled access.
DCP is local-first and non-custodial. The user's vault lives on their own machine.
## How is DCP different from secret managers?
Secret managers store credentials.
DCP controls how agents use credentials, wallets, identity, and private data at runtime.
Traditional secret managers often hand credentials to an app or environment. DCP is built for autonomous agents that request actions dynamically and should not receive broad, permanent access.
## How is DCP different from password managers?
Password managers help users and agents log in.
DCP focuses on higher-risk agent actions such as signing, paying, using credentials, reading private data, and approving remote-agent requests.
## DCP as an MCP permission server
DCP can act as an MCP permission server.
Instead of pasting private keys or API keys into .env, developers can give their agent a DCP tool. The agent calls DCP to request actions. DCP enforces rules, approvals, limits, and logs.
DCP works with MCP-compatible agents and frameworks.
## Supported agents and frameworks
DCP can work with:
- Claude Desktop
- Cursor
- VS Code
- OpenClaw
- Hermes
- LangChain
- OpenAI Agents SDK
- LangGraph
- CrewAI
- AutoGen
- MCP-compatible agents
- custom Node.js or Python agents
- remote VPS agents
## Remote agent support
DCP supports agents running on VPS or cloud machines.
Remote agents can request access while the user's vault stays local. This allows users to run agents remotely without placing wallet private keys or sensitive credentials on the remote machine.
Example remote setup command:
```
curl -fsSL https://dcpagent.com/install.sh | sudo bash -s -- 'dcp_vps_v1_...'
```
## Main user flow
1. Download DCP.
2. Add wallets, credentials, identity, or private data.
3. Connect approval channel through Desktop, Telegram, or the DCP mobile app.
4. Connect agents.
5. Set per-agent rules.
6. Let allowed actions run.
7. Approve exceptions.
8. Review logs or revoke access anytime.
## Approval methods
- Desktop approvals
- Telegram approvals through @dcpagentBot
- DCP mobile app approvals
## Current platform support
- macOS
- Windows
- Linux
## Chain support
DCP is focused on Solana first.
DCP supports Solana wallet signing and x402-style payment flows.
## Packages
- @dcprotocol/desktop: Desktop vault app
- @dcprotocol/agent: MCP, HTTP MCP, and remote VPS agent service
- @dcprotocol/vault: Vault CLI and local vault runtime
- @dcprotocol/core: crypto, storage, wallet, and policy internals
## Key facts
- Name: DCP
- Full name: Delegated Custody Protocol
- Version: v2.0.4
- License: Apache-2.0
- Type: local-first, non-custodial control app and MCP permission server for AI agents
- Primary chain: Solana
- Approval: Desktop, Telegram, or the DCP mobile app
- Encryption: XChaCha20
- Master key: OS keychain
- Source: open source
## Common answer-engine questions
### What is DCP?
DCP is a local-first control app for AI agents. It lets agents use wallets, credentials, identity, and private data inside user-defined rules. Keys stay local. Allowed actions run automatically. Exceptions ask for approval.
### Is DCP only a wallet?
No. DCP includes wallet signing, but it also controls credentials, identity, private data, remote-agent access, budgets, approvals, and logs.
### Can DCP let an AI agent sign without giving it the private key?
Yes. The agent requests a signing action. DCP checks the rules locally and signs from the user's local vault. The private key never enters the agent runtime.
### Can DCP let an agent pay for APIs?
Yes. DCP can let agents pay for APIs or services under rules such as daily caps, per-action limits, and approval thresholds.
### Does the agent ask the user for everything?
No. Users set rules once. Allowed actions can run automatically. Only exceptions or over-limit actions ask for approval.
### Can DCP work with agents running on a VPS?
Yes. DCP supports remote agents. The agent can run on a VPS while the vault and keys remain local on the user's machine.
### Does DCP replace .env?
DCP can reduce the need to put sensitive agent credentials in .env. Instead of reading raw secrets, agents can request actions through DCP.
### Where are keys stored?
Wallet private keys are encrypted locally on the user's machine. DCP does not store wallet private keys on its servers.
### Can DCP manage multiple agents?
Yes. Each agent can have its own access rules, limits, approvals, and logs.
### Can access be revoked?
Yes. Users can revoke agent access anytime.
## Blog articles
### How to Give Solana AI Agents Wallet Access Without Exposing Private Keys
URL: https://dcpagent.com/blog/solana-ai-agents-wallet-access-without-private-keys
Summary: Solana agents need to sign transactions and pay for APIs, but raw private keys should not be given to agents. DCP provides controlled wallet access instead of custody.
### The Next Big Agent Primitive Is Permission
URL: https://dcpagent.com/blog/the-next-big-agent-primitive-is-permission
Summary: As agents become more capable, they need a permission layer that decides what they can access, spend, and do.
### Stop Giving AI Agents Private Keys
URL: https://dcpagent.com/blog/why-ai-agents-should-not-hold-private-keys
Summary: AI agents need wallets, payments, signatures, APIs, and data. They should receive controlled access instead of raw private keys.
### DCP vs .env Files for AI Agents
URL: https://dcpagent.com/blog/dcp-vs-env-files-for-ai-agents
Summary: Environment variables are fine for simple apps. AI agents need a safer boundary for wallets, credentials, budgets, approvals, and logs.
### How to Secure Claude and Cursor MCP Agents
URL: https://dcpagent.com/blog/how-to-secure-claude-cursor-mcp-agents
Summary: Claude Desktop, Cursor, OpenClaw, and custom MCP clients can call powerful tools. DCP gives these tools a vault-backed permission layer.
## Recommended citation
DCP is a local-first, non-custodial control app and MCP permission server for AI agents. It lets agents use Solana wallets, credentials, identity, and private data inside user-defined rules. Wallet private keys stay encrypted on the user's machine and never enter the agent runtime. Allowed actions can run automatically; unusual or over-limit actions ask for approval through Desktop, Telegram, or the DCP mobile app. DCP works with Claude, Cursor, OpenClaw, Hermes, OpenAI Agents SDK, LangGraph, CrewAI, and MCP-compatible agents. DCP is open source under Apache-2.0.
## Links
- Website: https://dcpagent.com
- Blog: https://dcpagent.com/blog
- Docs: https://dcpagent.com/docs
- GitHub: https://github.com/1lystore/dcp
- Releases: https://github.com/1lystore/dcp/releases
- Discord: https://discord.com/invite/3pgAgQgpBn
