Skip to main content
Glama
AnB-MCP

AnB-MCP

AnB-MCP

An MCP server front-end for AnB that lets AI agents use secrets without ever seeing them.

anb-mcp quickstart

Headline guarantee: even a fully prompt-injected agent, calling every tool in every way, cannot extract a raw key. No tool returns a plaintext secret; reveal paths require a TTY that this server does not have, so alice refuses them.

Unlike a naive "secrets MCP" that hands the key to the model, here the agent gets placeholders and outcomes — the plaintext stays behind the anb-mcp → alice → Bob boundary.

How it works

Agent (untrusted) ──MCP/stdio──► anb-mcp ──exec──► alice ──mTLS──► Bob ──► master key
                                  (this repo)      (AnB client)   (AnB KMS daemon)

anb-mcp runs as a dedicated, narrowly-scoped AnB identity (not your operator CLI identity), so a compromised agent's blast radius is limited to what Bob authorizes for that identity.

Related MCP server: ChatGPT Gateway MCP Nyan

Tools

Tool

Does

Returns

anb_list

List secret keys this identity may reference

names + metadata, no values

anb_exec

Run an operator-allowlisted command with secrets injected into the child's env

exit code + redacted stdout/stderr

anb_status

Health / authz self-check

Bob reachability, identity, authorized prefixes, rule count

anb_redact

Scrub text — secret values + high-entropy tokens → <agent-vault:key>

redacted text

anb_render_to_file

Render a placeholder template, write a 0600 file under the render dir

the path, never the content

Never exposed: any reveal / get-plaintext / shell tool.

Prerequisites

This is a thin front-end; it depends on AnB. For v0.1 you need:

  1. A working alice + bob (AnB) on the host.

  2. A dedicated MCP identity enrolled with Bob, scoped to only the key prefixes the agent should use. Point the server at it via ANB_MCP_ALICE_DIR (default ~/.anb/alice-mcp). Do not reuse your operator identity.

  3. Exec allowlist with scope tags — alice's exec rules carry a 4th scope column; only rules tagged mcp apply to this surface (default-deny). Tag a rule for the agent by appending mcp (e.g. ^/opt/.../curl ...$\tOPENAI_KEY\t# call\tmcp). (Requires AnB with alice exec --surface, alice redact, and alice status --json — all shipped.)

Build

go mod tidy
go build -o anb-mcp .

Register with Claude Code

claude mcp add -s user -e ANB_MCP_ALICE_DIR=$HOME/.anb/alice-mcp \
  anb -- /path/to/anb-mcp

Or in ~/.claude.json under mcpServers:

{
  "mcpServers": {
    "anb": {
      "command": "/path/to/anb-mcp",
      "env": { "ANB_MCP_ALICE_DIR": "/Users/you/.anb/alice-mcp" }
    }
  }
}

Tools surface as mcp__anb__anb_list, mcp__anb__anb_exec, mcp__anb__anb_status.

Status

v0.1 — done, verified end-to-end (and by a real agent). All three tools work against a live Bob: anb_status returns real KMS state; anb_exec runs allowlisted commands and denies the rest; and a secret injected via --env <agent-vault:key> is used by the child process while the caller receives only the redacted placeholder — the plaintext never reaches the agent. Confirmed both by go-sdk-client invariant tests (test/) and by an independent Claude Code session calling the tools over MCP. See CHANGELOG.md.

Roadmap: see PLAN.md. (v0.2 adds anb_render_to_file + a dedicated anb_redact tool; v0.3 lowers per-call latency and adds per-agent ephemeral, short-TTL scoped credentials — while keeping alice as a separate process, so the no-reveal guarantee stays structural, not a code-discipline promise.)

License

MIT

Available Tools

5 tools
anb_execA

Run an operator-allowlisted command with named secrets injected into the child process's environment. SIDE EFFECT: spawns a real subprocess — default-deny, only commands matching a scope=mcp allowlist rule run, everything else is refused. NOT idempotent (effects depend on the command). Requires an enrolled identity and a reachable, unlocked Bob to resolve agent-vault:key placeholders. Returns the exit code plus REDACTED stdout/stderr; the raw secret is never returned even if the child prints it. A denied or failed command returns a non-zero exit_code with the reason in stderr_redacted.

Inputs: command (required), env, args. Returns: exit_code, stderr_redacted, stdout_redacted

anb_listA

List the secret key names and metadata this identity may reference — never the values. Read-only and idempotent, no side effects. Requires an enrolled identity and a reachable, unlocked Bob KMS daemon; errors if Bob is unreachable/locked or the identity is unauthorized. Use this to discover which agent-vault:key names exist before referencing them in anb_exec or anb_render_to_file.

Returns: keys

anb_redactA

Scrub text: replace known secret values and high-entropy tokens with agent-vault:key placeholders. Read-only, idempotent, no side effects, and needs no Bob connection. Fail-safe by design — may over-redact, never under-redact. Use before logging or returning any text that might contain a secret.

Inputs: text (required). Returns: redacted

anb_render_to_fileA

Render a template containing agent-vault:key placeholders and write the resolved file (mode 0600) under the confined render dir. SIDE EFFECT: writes a file to disk (overwrites if the path exists). Requires an enrolled identity and a reachable, unlocked Bob to resolve placeholders. Returns the written path, NEVER the resolved content — the caller never sees the secret values. out_path is relative to the render dir; absolute paths and .. traversal are rejected.

Inputs: out_path (required), template (required). Returns: path, written

anb_statusA

Read-only self-check: reports enrollment, client-cert presence, Bob reachability and unlock state, identity, server name, and idle-TTL. No side effects, idempotent, returns NO secret values. Call this first to confirm the vault is ready before anb_list/anb_exec/anb_render_to_file; on failure the error field explains why (e.g. Bob unreachable or locked).

Returns: error, bob_addr, enrolled, identity, +5 more

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.7/5.0

    Scored across 5 tools

    Disambiguation5/5

    Each tool has a clearly distinct purpose: executing commands with secret injection, listing secret keys, redacting secrets from text, rendering templates to disk, and checking server status. There is no overlap or ambiguity between them.

    Naming Consistency5/5

    All tool names follow a consistent 'anb_<verb>' pattern, with verbs like exec, list, redact, render_to_file, and status. This makes the tool set predictable and easy to navigate.

    Tool Count5/5

    With exactly 5 tools, the server is well-scoped for its purpose—providing secure vault-integrated operations without unnecessary complexity. The count feels appropriate for the domain.

    Completeness5/5

    The tools cover the full lifecycle needed for a secret-aware MCP server: status checking, secret listing, command execution, template rendering, and output redaction. While there is no tool for managing secrets themselves, that is intentional for a client-side tool, and the set is complete for its intended use.

    Maintenance

    ActivityInactive
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables users to define and run MCP tools using declarative YAML configs with built-in trust enforcement, credential brokering, and tamper-evident audit logging.
      14
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables agent clients to safely connect to tools and execution resources through MCP with authorization, approvals, audit, chat-context isolation, SSH/Docker access, and long-running command session tracking.
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables MCP clients to run declarative agents and DAG workflows as plain tools, with parallel nodes, review loops, and per-run least-privilege sandboxing.
      5 npm
      MIT
    • A
      license
      Not graded
      quality
      C
      maintenance
      Provides an MCP stdio interface for AI agents to propose tool calls that are validated against task policy, held for human approval when required, executed in a sandboxed Linux environment, and recorded in an audit log.
      MIT