abadge

Your agents need credentials. You need control.

abadge is a credential control plane for AI agents. Store secrets, permission exactly which agents can use them, choose how they're delivered, and audit every access -- all from one place.

---

The problem: Agents are shipping to production. They call APIs, sign into services, and act on behalf of users. But credentials are still hardcoded in config, over-shared in environment variables, or buried in vaults with no per-agent scoping. When something goes wrong, there's no trace.

The fix: abadge puts a policy-checked, audited gateway between every agent and every credential.

---

How it works

Store → Permission → Deliver → Audit

1. Store credentials with zero-knowledge or server-managed encryption
2. Permission each agent for exactly the credentials and capabilities it needs
3. Deliver secrets through env injection, file mounts, or direct reveal -- LLMs never see raw secrets
4. Audit every access attempt in an append-only log

Surfaces

	For	How
Dashboard	Operators	Web UI for managing credentials, agents, and permissions
CLI	Developers	abadge run injects secrets into any command
MCP Server	AI Agents	Tools that use secrets without exposing them to the model
SDK	Integrators	TypeScript client for programmatic access
API	Everything	tRPC control plane on Cloudflare Workers

Quick start

# Install
curl -fsSL https://raw.githubusercontent.com/punitarani/abadge/main/install.sh | bash

# Login
abadge login

# Store a secret
abadge item create --label "prod-db" --kind login --storage-mode server_managed

# Register an agent and grant access
abadge agent register -n "deploy-bot"
abadge permission create --agent-id <id> --item-id <id> --capability mount_env

# Run with the secret injected
abadge run --item <id> --env-var DB_PASSWORD -- ./deploy.sh

Zero-knowledge option

For maximum security, choose zero-knowledge storage. Your master password derives the encryption key locally -- the server stores only ciphertext and can never decrypt your secrets, even in a full breach.

Security highlights

• AES-256-GCM (server-managed) and XChaCha20-Poly1305 (zero-knowledge) encryption at rest
• API keys and session tokens hashed before storage, shown once
• Ed25519 signed sessions with 15-minute TTL for agents
• Per-agent, per-credential, per-capability permissions with optional expiry
• MCP tools redact secrets from LLM-visible output
• Append-only audit log for every access attempt (allowed and denied)
• Rate limiting, secure headers, CORS, CSRF protection

Architecture

flowchart TD
  subgraph Surfaces
    Dashboard["Dashboard"]
    CLI["CLI"]
    MCP["MCP Server"]
    Agent["Remote Agent"]
  end

  subgraph Local["Local Runtime"]
    Daemon["Vault Daemon\n(ZK crypto)"]
  end

  subgraph Edge["Cloudflare Edge"]
    API["API\n(Hono + tRPC)"]
  end

  DB[(PostgreSQL)]

  Dashboard --> API
  CLI --> API
  CLI --> Daemon
  MCP --> API
  MCP --> Daemon
  Agent --> API
  API --> DB

Loading

Documentation

Document	Description
Overview	What abadge is, who it's for, and how it works
Workflows & Flows	Mermaid diagrams for every workflow across all surfaces
Entities & Data Model	Database schema, entity relationships, lifecycle
Security Model	Encryption, auth, authorization, threats, and trust boundaries
Architecture	System design and package structure
API Spec	tRPC procedures reference
CLI Spec	Command reference
MCP Spec	MCP tool reference
SDK Spec	TypeScript client reference
Crypto Spec	Cryptographic envelope and key hierarchy
Threat Model	Trust boundaries and breach analysis
Development	Setup, commands, and contributing

Tech stack

Hono + tRPC on Cloudflare Workers, Next.js dashboard, PostgreSQL via Hyperdrive, Drizzle ORM, Better Auth, Zod, Turborepo + Bun.

Development

bun install          # Install dependencies
bun run docker:up    # Start local Postgres
doppler setup        # Configure secrets
bun run db:push      # Push schema
bun run dev          # Start dev servers

Status

Active development. Building the credential control plane for the agent era.