server

---

Try It Now (2 minutes)

See AgentCred verification in action (GitHub auth required):

# Initialize identity (once) — this starts the GitHub OAuth device flow
npx @agentcred-ai/cli init

# Sign a message
echo "Hello from a verified AI agent!" | npx @agentcred-ai/cli sign --agent demo-bot > envelope.json

# Verify it
npx @agentcred-ai/cli verify < envelope.json

TIP: To skip the OAuth flow, pass a token:

npx @agentcred-ai/cli init --token ghp_your_token

---

The Problem

AI agents are everywhere — writing code, posting comments, sending emails. But they're anonymous.

Platform	The Problem
Reddit	Can't tell who's a bot
YouTube	Can't trace comment sources
Your inbox	Can't verify if "AI assistant" is trustworthy

When an agent makes a mistake or spreads misinformation, there's no accountability.

The Solution

AgentCred gives every AI agent a verifiable identity tied to a real human:

Agent Output → Cryptographic Signature → GitHub Identity → ✓ @username

• Ed25519 signatures: Same crypto that secures billions of SSH connections
• GitHub identity: Developer trust you already recognize
• Open standard: Not controlled by any single company

---

Who Is This For?

You Are...	Your Pain Point	AgentCred Solves It
AI Agent Developer	"My bot looks like every other spam bot"	Signatures establish trust and reputation
Platform Operator	"Can't trace AI-generated content"	Verification API for filtering and moderation
Enterprise	"Need audit trail for AI actions"	Every action attributed to responsible human
End User	"Is this bot trustworthy?"	Browser badge shows verification instantly

---

Installation

Pick your IDE or tool to get started:

Config file location:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "agentcred": {
      "command": "npx",
      "args": ["-y", "@agentcred-ai/mcp-server"]
    }
  }
}

Restart Claude Desktop. Done!

Config file location: ~/.cursor/mcp.json

{
  "mcpServers": {
    "agentcred": {
      "command": "npx",
      "args": ["-y", "@agentcred-ai/mcp-server"]
    }
  }
}

Restart Cursor. Done!

Config file location: .vscode/mcp.json (workspace) or ~/.vscode/mcp.json (global)

{
  "mcpServers": {
    "agentcred": {
      "command": "npx",
      "args": ["-y", "@agentcred-ai/mcp-server"]
    }
  }
}

Restart VS Code. Done!

Config file location: ~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "agentcred": {
      "command": "npx",
      "args": ["-y", "@agentcred-ai/mcp-server"]
    }
  }
}

Restart Windsurf. Done!

Config file location: ~/.config/zed/settings.json

{
  "context_servers": {
    "agentcred": {
      "command": "npx",
      "args": ["-y", "@agentcred-ai/mcp-server"]
    }
  }
}

Restart Zed. Done!

# Add to your shell config (~/.bashrc, ~/.zshrc, etc.)
export MCP_SERVERS='{"agentcred":{"command":"npx","args":["-y","@agentcred-ai/mcp-server"]}}'

# Or run inline:
MCP_SERVERS='{"agentcred":{"command":"npx","args":["-y","@agentcred-ai/mcp-server"]}}' claude

Done!

# Install globally
npm install -g @agentcred-ai/cli

# Or use with npx (no install)
npx @agentcred-ai/cli --help

npm install agentcred

import { sign, verify } from 'agentcred'

---

Authentication

AgentCred links your AI agent to your GitHub identity. Two methods are supported:

Option 1: OAuth Device Flow (Default)

When you run init, your browser opens automatically for GitHub login:

npx @agentcred-ai/cli init

What happens:

1. CLI shows a one-time code (e.g., ABCD-1234)
2. Browser opens to github.com/login/device
3. Enter the code and authorize
4. Ed25519 keypair generated and registered

No secrets to manage! OAuth tokens are handled automatically.

Option 2: Personal Access Token (PAT)

For CI/CD, headless servers, or if OAuth doesn't work:

# Via flag
npx @agentcred-ai/cli init --token ghp_your_token

# Via environment variable
export GITHUB_TOKEN=ghp_your_token
npx @agentcred-ai/cli init

For MCP Server (Claude Desktop, Cursor, etc.):

{
  "mcpServers": {
    "agentcred": {
      "command": "npx",
      "args": ["-y", "@agentcred-ai/mcp-server"],
      "env": {
        "GITHUB_TOKEN": "ghp_your_token",
        "GITHUB_USERNAME": "your-github-username"
      }
    }
  }
}

Create a GitHub PAT here (only read:user scope needed)

---

Quick Start

For AI Assistants (Claude, Cursor, etc.)

1. Add AgentCred to your MCP config (see Installation)
2. Restart your AI assistant
3. Ask: "Initialize AgentCred and sign this message: Hello world"

That's it! Your agent now has a verifiable identity.

For CLI

# Step 1: Initialize (opens browser for GitHub OAuth)
npx @agentcred-ai/cli init

# Step 2: Sign anything
echo "Hello world" | npx @agentcred-ai/cli sign --agent my-bot > envelope.json

# Step 3: Verify
npx @agentcred-ai/cli verify < envelope.json

# Check your identity
npx @agentcred-ai/cli whoami

Using PAT instead? See Authentication.

For Web Content (Invisible Signatures)

For blogs, social media, and web pages, use signWithHTML() which embeds signatures invisibly:

import { signWithHTML } from 'agentcred'

const html = await signWithHTML("Hello world", identity, { agent: "my-bot" })
// Returns: <span data-agentcred="...">Hello world</span>
// Verification tools can detect and verify this automatically.

---

How It Works

┌─────────────────────────────────────────────────────────────────────────────┐
│                                                                             │
│  1. AUTHENTICATE      2. GENERATE        3. SIGN          4. VERIFY        │
│  ┌─────────────┐     ┌─────────────┐   ┌─────────────┐  ┌─────────────┐    │
│  │   GitHub    │     │  Ed25519    │   │   Agent     │  │   Anyone    │    │
│  │   OAuth     │────▶│  Keypair    │──▶│   Signs     │─▶│  Verifies   │    │
│  │   Login     │     │  Generated  │   │   Output    │  │  ✓ Badge    │    │
│  └─────────────┘     └─────────────┘   └─────────────┘  └─────────────┘    │
│        │                   │                 │                │            │
│        ▼                   ▼                 ▼                ▼            │
│   "I am @alice"      Public key        JWS signature    "✓ @alice"        │
│                      registered        + SHA-256 hash    verified          │
│                                                                             │
└─────────────────────────────────────────────────────────────────────────────┘

Step	What Happens	Where
1. Authenticate	Verify GitHub identity via OAuth or PAT	Browser / Terminal
2. Generate Keys	Create Ed25519 keypair (public + private)	Your machine (~/.agentcred/)
3. Register	Link public key to GitHub username	AgentCred API
4. Sign	Agent creates JWS with content hash	Your agent
5. Verify	Check signature against registered key	Anyone, anywhere

Output format (AgentCred Envelope):

{
  "agentcred": {
    "v": "1.0",
    "jws": "eyJhbGciOiJFZERTQSIs...",
    "github": "username",
    "agent": "bot-name"
  },
  "content": "The signed content"
}

---

Real-World Use Cases

1. Verified Bot Comments

Problem: Your helpful AI bot on Reddit/Discord is indistinguishable from spam.

Solution: Sign every message. Users see ✓ @yourname badge.

const response = await myBot.generateResponse(userQuestion)
const signed = await sign(response, identity, { agent: "support-bot" })
await postToDiscord(signed)  // Recipients can verify

2. Enterprise Audit Trail

Problem: Compliance requires knowing who's responsible for AI-generated reports.

Solution: Every AI action is cryptographically attributed.

# CrewAI agent with AgentCred
analyst = Agent(
    role="Financial Analyst",
    tools=mcp_adapter.get_tools(),  # AgentCred signing
)
# All outputs signed: "Report generated by @alice's agent at 2026-02-03"

3. Auto-Sign All AI Outputs

Problem: You want every response from your AI to be signed automatically.

Solution: Middleware wraps all outputs.

// Vercel AI SDK - 3 lines to sign everything
const signedModel = wrapLanguageModel({
  model: openai('gpt-4o'),
  middleware: createAgentCredMiddleware({ github: 'you', privateKey }),
})
// Every response is now signed!

4. Content Authenticity

Problem: Team can't verify if AI analysis is legitimate.

Solution: Share signed envelope, anyone can verify.

# Alice signs her AI's analysis
echo "Q4 revenue up 15%" | npx @agentcred-ai/cli sign > analysis.json

# Bob verifies it came from Alice
npx @agentcred-ai/cli verify < analysis.json
# ✓ Verified: @alice (analyst-bot) at 2026-02-03T18:36:09Z

---

Framework Integrations

Framework	Language	Effort
Claude Desktop	—	Config only	Setup
Vercel AI SDK	TypeScript	3 lines	Example
Mastra	TypeScript	3 lines	Example
LangChain	Python	10 lines	Guide
CrewAI	Python	10 lines	Guide
Any CLI	Any	Shell out	npx @agentcred-ai/cli sign

import { createAgentCredMiddleware } from '@agentcred-ai/vercel'
import { wrapLanguageModel } from 'ai'

const signedModel = wrapLanguageModel({
  model: openai('gpt-4o'),
  middleware: createAgentCredMiddleware({ github: 'you', privateKey }),
})

import os

from langchain_mcp_adapters.client import MultiServerMCPClient
from langchain_openai import ChatOpenAI
from langgraph.prebuilt import create_react_agent

client = MultiServerMCPClient({
    "agentcred": {
        "command": "npx",
        "args": ["-y", "@agentcred-ai/mcp-server"],
        "transport": "stdio",
        "env": {
            "GITHUB_TOKEN": os.environ["GITHUB_TOKEN"],
            "GITHUB_USERNAME": os.environ["GITHUB_USERNAME"],
        },
    }
})
tools = await client.get_tools()  # agentcred_sign, agentcred_verify
agent = create_react_agent(ChatOpenAI(model="gpt-4o"), tools)

---

Packages

Package	Description
agentcred	Core SDK
@agentcred-ai/sdk	Full SDK with all exports
@agentcred-ai/mcp-server	MCP server for Claude
@agentcred-ai/cli	Command-line tool
@agentcred-ai/vercel	Vercel AI SDK middleware
@agentcred-ai/mastra	Mastra tool wrapper

---

Documentation

• guide.md — Beginner-friendly walkthrough
• examples/ — Working code examples

---

Roadmap

• Web verification UI at agentcred.dev/verify
• Embeddable "Verified Agent" badge component
• Complete OpenClaw & Moltbook integrations

Want to see another framework supported? Open an issue or contribute!

---

Contributing

Contributions welcome! See CONTRIBUTING.md.

License

This project is licensed under the MIT License - see the LICENSE file for details.

---