MCP overview - abadge

The abadge MCP server exposes item-aware tools to AI runtimes via Model Context Protocol. It runs as a subprocess over stdio and is consumed by clients like Claude Desktop, MCP-aware IDEs, and custom agent runtimes. The model never receives raw secret material. Tools either operate on metadata or inject secrets into subprocesses behind a redaction boundary.

Transport

Property	Value
Protocol	MCP
Transport	stdio
Server name	`abadge-mcp`

Quick reference

Tool	Purpose
`list_items`	Item metadata only
`use_secret`	Run a command with a secret injected as an env var. Subprocess output is never returned to the model (only exit code, duration, line counts).
`mount_secret`	Mount a secret as a `0600` temp file. Returns an opaque `mountId`. Auto-deletes after 5 minutes.
`release_mount`	Release a mounted secret early.
`get_audit`	Fetch recent audit entries.

Auth at a glance

Ed25519 keypair sessions — the same auth the SDK’s AbadgeAgentClient uses.

export ABADGE_API_URL=http://localhost:8787
export ABADGE_AGENT_ID=agent_...
export ABADGE_PRIVATE_KEY_PATH=~/.abadge/agents/<agent-id>.ed25519.jwk

The MCP server runs an Ed25519 challenge / session exchange on startup and refreshes the abs_... token before it expires. No long-lived secret ever sits on disk. Keypair sessions are the only auth method — there are no agent API keys. See Authentication for the full configuration matrix, MCP-specific overrides, and the ~/.abadge/config.json fallback.

Where to go next

Installation

Run via bun run mcp or the compiled binary.

Authentication

Env vars, config fallback, and the keypair handshake.

Claude Desktop

The mcpServers JSON snippet to drop into your client config.

Security model

What the model sees vs doesn’t, and the redaction boundary.

​Transport

​Quick reference

​Auth at a glance

​Where to go next