AgentFox

Find, govern and protect every AI agent you run. Guardrails that don't block your real customers.

Open source. Attack-test your own agent in about three minutes, then watch the same attacks get stopped. Works with the OpenAI Agents SDK, LangGraph, LangChain, CrewAI, AutoGen, OpenAI, Anthropic, LiteLLM, MCP servers, Claude Code and Codex.

▶ Try it in the browser, no account  ·  Docs  ·  Benchmarks  ·  Discussions

If AgentFox is useful to you, a ⭐ helps other people find it.

What it does

The part that does not depend on catching the attack: a transfer whose recipient came out of a retrieved document is refused because of where the value came from, not because a detector recognised the payload.

Quickstart

Python 3.11 or newer.

Attack your own agent (about three minutes)

1. Sign in; signing in creates your workspace. Press Create a key.

2. In your agent's repository, with its dependencies installed, run the command the page shows:

   pip install -U agentfox
   agentfox init --key nom_api_… --server https://…
   

   init reads your code without running it and reports the agents and tools it found. Leave it running: the attack test runs your agent on your machine, with every tool faked so nothing real happens.

3. Press Add N agents to AgentFox. The attack test starts by itself and shows, per kind of attack, how many got through.

4. Press Add and test again. Basic guardrails go on and the same attacks run again, so you see the before and after side by side, including how many normal requests were still answered.

5. Add the one line init prints (agentfox.auto("<agent>")) and run your app. Its calls show up on the agent's Activity tab.

From the command line only

pip install agentfox
agentfox scan                    # names every agent it found, ends with: agentfox up
agentfox up                      # registers them, declares their tools, adds agentfox.auto()
python main.py                   # run your app as usual
agentfox findings                # what it saw, and what it would have stopped
agentfox test redteam <agent>    # attack-test it

up uses the names the code gives your agents, the same names scan printed, and adds agentfox.auto("<agent>") to the file that starts one (a pyproject.toml script, the graph langgraph.json loads, or the file that runs it, src/ included), for Python apps on LangGraph, LangChain, CrewAI, AutoGen, pydantic-ai and the OpenAI Agents SDK. The name is the agent that file runs (a crew or swarm takes its project's name), shown before anything changes so you can edit it. In a folder of separate examples it sets up one, with only that agent's tools, and says how to pick another (agentfox up --agent <name>). It starts in watch mode and changes nothing else by default: coding-agent hooks, the GitHub workflow and the local API are --with coding-agents, --with ci and --with view. Everything stays on your machine; a pip install has no web dashboard (agentfox findings is the terminal view), which comes with AgentFox Cloud or the self-hosted stack.

The prompt-injection classifier (PIGuard) is an extra, not part of pip install agentfox:

pip install 'agentfox[classifiers]'
agentfox admin detectors pull injection.classifier   # downloads the model weights once

Without it, injection checks use the pattern-based detector only, alongside the PII, secrets and hidden-Unicode checks and tool containment, all of which work on a plain install. agentfox doctor says which detectors run on your machine, and auto() says once that the classifier is skipped rather than on every run. The Docker image ships the extra and the weights.

One key everywhere

export AGENTFOX_URL=https://…
export AGENTFOX_API_KEY=nom_api_…

With these set (or .agentfox/agentfox.env, or agentfox.auto(server=…, key=…)), every check agentfox.auto() makes goes to your account's gateway and the calls appear on the dashboard. Without a key it checks and records locally. Every OpenAI, Anthropic, LiteLLM and LangChain call in the process is checked, and every tool call in a response is checked before your code can run it; a refused one raises agentfox.Blocked.

Other ways in

• Check your machine (about 2 seconds): agentfox scan machine reads the MCP configs and skills that Claude Code, Cursor, Claude Desktop, VS Code, Windsurf, Gemini CLI and Codex load. Nothing is started, written or sent. Without installing: uvx agentfox scan machine.

• Your coding agent: in a project where you use Claude Code or Codex, agentfox up --with coding-agents writes the hooks, grants the agent its built-in tools and starts in watch mode, except shapes such as rm -rf ~/, which are refused from the start. In a repository set up for a coding agent with no agent of its own, agentfox up does this by default. The hooks call AgentFox by absolute path, so they work from a virtualenv install too.

• Behind a LiteLLM proxy: one block in the proxy's config.yaml checks every agent behind it (guide).

• Look around first: agentfox init && agentfox demo. init creates a SQLite database and loads 44 controls and the policy packs in about a second; demo runs a thirteen-step walkthrough in about five, in a database of its own (~/.agentfox/demo.db), so your project's agentfox findings never shows the demo's made-up agents. Both are offline: no API key, no downloaded weights, no network egress.

• Scan this directory without installing: curl -fsSL https://raw.githubusercontent.com/architsharm/agentfox/main/scripts/quickscan.sh | bash installs the published release into a throwaway virtualenv (| bash -s -- --main for the current main).

• On every pull request: the GitHub Action annotates lethal trifectas, ungoverned model calls and poisoned skills, and fails the check on the kinds you list in fail-on:

  - uses: actions/checkout@v4
  - uses: architsharm/agentfox@main
    with:
      fail-on: trifecta,skills
  

Contain a tool call

Any language can ask the gateway about one call before running it:

agentfox serve                              # gateway + control-plane API on 127.0.0.1:8080
agentfox admin users create you@example.com --token   # first operator + API token, shown once

curl -s -X POST http://localhost:8080/v1/guard/tool_call \
  -H "Authorization: Bearer $AGENTFOX_API_KEY" -H "Content-Type: application/json" \
  -d '{"agent":"payments-ops","tool":"payments.transfer",
       "arguments":{"amount":250,"currency":"USD","to":"acct_991"},
       "provenance":{"to":"tool_result","amount":"user"},
       "intent":"refund a duplicate charge"}' | jq '{verdict, approval_id}'

{ "verdict": "escalate", "approval_id": "apr_01m376q43zby33tbsp" }

The recipient came from a tool result, so the transfer waits for a human. Flip "to" to "user" and the same call returns allow. Once a person approves it (agentfox permit approvals approve apr_…), the same call with "approval_id" added runs, once. Pointed at /v1 as an OpenAI or Anthropic base URL, the gateway proxies model calls too, and a held one returns HTTP 428 with its approval_id. LangGraph, MCP, Claude Code hooks, the SDK and the full command set are in the website guides and the product tour.

Tracking without anyone running a command

A scan is a snapshot. Connect a GitHub repository, a hosted API's OpenAPI document or an MCP server once and AgentFox keeps re-checking it: every few hours, and on every push once the GitHub webhook is registered. What changed since the last run becomes a finding, and the finding closes itself when the condition clears. Setup and every finding type: Monitor connected sources.

What we measured

The detector comparison behind these numbers is peer reviewed: No Tier Wins Everywhere, accepted at WAITI 2026, a workshop of ACSAC 2026 (artifact).

We do not claim adversarial robustness, and we do not believe anyone can. The Attacker Moves Second (Nasr, Carlini, Schulhoff et al., 2025) reports over 90% attack success against twelve published defences once the attacker adapts. So the number we lead with does not depend on catching anything: both rows below were measured with every detector switched off.

The cost is benign utility: when provenance is inferred, a legitimate action that copies a value out of a tool output looks the same as an attack, and it is escalated.

Optionally, detection gets better. The judgment tiers are off by default; with them on, agentfox catches 160/165 of the injection payloads that defeated our own pattern detectors, at 94.7% precision against the NotInject over-defense set. What each tier is worth, and what it costs.

Where we are still improving

Detection is the layer we trust least, so we publish its numbers.

• The default injection stack (heuristic plus the PIGuard classifier) has a held-out recall of 68.3%, at 100% precision (REPORT.md). It needs agentfox[classifiers] and the model weights, which the Docker image ships; without them the heuristic runs alone, and its held-out injection recall is 16.7%; the startup warning and agentfox doctor say so.
• On NotInject, 339 benign prompts built to trip keyword guardrails, the classifier flags 11.5% (39/339) at the level an enforcing policy acts on (PIGuard alone), and 41% (140/339) counting the advisory backstop hits that observe mode records.
• An adaptive attacker that reads our verdict and retries gets 74% of the attacks we catch through within 50 attempts, measured against the heuristic-only stack that was the default until October 2026.
• The deterministic answerability classifier abstains on 57/676 contested questions; with a judgment tier that becomes 572/676, at the cost of over-refusal rising from 0.75% to 6.8%.
• Against a real, independently installed llm-guard on indirect injection via tool output, it is more precise than us: 81.8% against our 66.7% on the same 20 cases, though of the 10 attacks among them we catch all 10 and it catches 9. That run predates October 2026, while the backstop model's hits still counted toward a block (agent_security).
• On AgentDojo, per-argument taint misses 61 of 702 attacker write calls; session-level taint misses none of them and escalates three in four benign tasks.
• The default stack reaches 86.4% and 98.4% recall on two independent datasets (SPML and yanismiraoui), but on long prompts (around 2,000 characters and up) the classifier mostly times out within its 250 ms budget and the request falls back to the heuristic.

Treat every detection number as a speed bump that raises attacker cost, never as a defence.

Not built yet

• It is only as good as your declarations. A destructive tool declared read is treated as read. agentfox doctor grades this; agentfox scan finds undeclared tools.
• You declare the estate yourself. No Okta, no DataHub; principals and grants live in AgentFox.
• Compliance mappings are DRAFT, not reviewed by counsel, and labelled so in every export.
• No live IdP or SSO. Multi-tenancy is enforced at the session within one deployment, and only text is inspected.

Live per-capability status, computed by probe: docs/status.md.

Documentation

Questions and ideas go to Discussions; bugs to Issues. New contributors: start with a good first issue.

Licence

The open core is Apache-2.0: the SDKs, the gateway and CLI, enforcement and containment, every detector and validator, custom rules, policy packs, the benchmarks and the dashboard. It runs on its own, with no licence key and no limit on agents.

AgentFox Enterprise (ee/) adds what an organisation needs to govern many agents and prove it: compliance status against EU AI Act, ISO 42001, SOC 2 and NIST AI RMF, auditor evidence packages, fleet monitoring with alerts, and checks against your own access control. It is source-available under the Business Source License 1.1: free in production for one workspace with up to four agents, and each release becomes Apache-2.0 three years after it ships. It is included on AgentFox Cloud.