# Check an agent's tool call before it runs

> Check an agent's tool call before it runs: 0 companies fully do it, 8 with a limit, 20 carry the tag. Every listing in Guardrails & safety filters with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/guardrails.gate-tool-calls/
- Markdown: https://www.anchorterminal.com/niches/guardrails.gate-tool-calls/index.md (~1,600 tokens)
- Slim: https://www.anchorterminal.com/niches/guardrails.gate-tool-calls/index.min.md (~380 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/guardrails.gate-tool-calls/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

No company's docs show a full fit yet. The listings below do the job with a limit, or not at all.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 0 | 8 | 20 | 9 | 5 |

## What counts as done

The tool name and arguments are screened before execution, the verdict is allow, deny or ask a person with a reason, secrets and personal data in arguments are caught, and a turn that holds only tool calls is not skipped.

How agents ask for it: stop the agent running dangerous shell commands; check this tool call before executing it; block tool calls that leak secrets or customer data

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **AgentGate (Mohit Singh (individual))**: POST /v1/assess scores tool name and arguments before execution and returns allow, deny or ask with reason_codes, and on_model_error defaults to ask; but secret or PII detection in arguments is not a head, and the classifier has no published accuracy.
- **Cisco AI Defense Inspection API (Cisco Systems, Inc.)**: MCP inspection covers call_tool and list_tools messages before they run and returns is_safe with the triggered rules, but there is no ask-a-person verdict.
- **Granite Guardian (International Business Machines Corporation)**: available_tools and the function calling hallucination criterion check a call against the declared tools, and custom criteria can state a rule, but the verdict is yes or no with no ask outcome and secrets are not a pre-baked target.
- **Lakera Guard (Check Point AI Guardrails) (Check Point Software Technologies Ltd.)**: tool_calls are screened as the agent's actions with a Tool Allow/Deny List and a beta Dangerous Deviation detector, but the verdict is flagged or not, with no ask-a-person outcome.
- **LlamaFirewall (Meta Platforms, Inc.)**: AlignmentCheck audits the trace and assistant tool_calls against the user's goal, and ScanDecision has allow, block and human_in_the_loop_required; but it calls Together AI by default and nothing targets secrets in arguments.
- **NVIDIA NeMo Guardrails (NVIDIA Corporation)**: Tool call validation blocks undeclared tools and schema-violating arguments, fails closed and names the reason, but it is structural only: no secret or PII screening of arguments and no ask-a-person verdict.
- **OpenAI Guardrails (OpenAI (as named in the LICENSE copyright line and the PyPI author field))**: Before tool calls run, the check tests whether the requested function fits the user's goal and trips a block, but there is no ask-a-person verdict and secrets in arguments are not named as a target.
- **Prisma AIRS AI Runtime Security API (Palo Alto Networks, Inc.)**: tool_event takes the raw JSON input of an MCP or function call and tool_detected reports findings with an allow or block action, but there is no ask-a-person verdict.

## Couldn't check

The docs were closed to us, or silent. These carry no mark either way.

- **Guardrails AI (Guardrails AI, Inc.)**: Issue 1601 on validating tool calls with on="tools:..." is closed, but no shipped docs or release note confirming it were read.

- not this job: Amazon Bedrock Guardrails (Amazon Web Services, Inc.): The docs state that in tool-use workloads tool-call arguments, tool results and tool definitions are not evaluated, and acc-18 saw tool-call-only turns never sent; there is no allow, deny or ask verdict on an action.
- not this job: Azure AI Content Safety (Prompt Shields) (Microsoft Corporation): Prompt Shields scan user input and, in preview for agents, tool responses; nothing read screens tool-call names and arguments or returns an allow, deny or ask verdict.
- not this job: Google Cloud Model Armor (Google LLC): The sanitize calls take a user prompt or a model response; the overview names no tool-call field, tool definition screening or allow, deny or ask verdict.
- not this job: Llama Guard 4 (Meta Platforms, Inc.): S14 Code Interpreter Abuse covers text only; there is no tool-call input or allow, deny or ask verdict.
- not this job: Presidio (Data Privacy Stack (community organisation, no legal entity stated)): It can find PII in argument text a caller passes, but returns entities, not an allow, deny or ask verdict on a tool call.
- not this job: Mistral Moderation API (Mistral AI (RCS Paris 952 418 325)): Moderation classifies text or the last conversation turn; no tool-call argument screening or allow, deny or ask verdict is documented.
- not this job: OpenAI Moderation API (OpenAI OpCo, LLC): Classifies text and images; no tool-call input or allow, deny or ask verdict.
- not this job: ShieldGemma (Google LLC (RDAP registrant of gemma.ai; the Gemma Terms of Use are Google's and the cards name Google as author)): Takes a prompt or a prompt and response; no tool-call input or verdict.

## What stops agents finishing

- guardrail screens message text only, so tool-call-only turns pass unchecked
- ask verdict needs a human with no channel to reach them

## Around it

Before: inference.tool-loop. After: human-in-the-loop.approve-action, code-sandboxes.run-generated-code.

## Where this comes from

Weight 9 of 100 in Guardrails & safety filters: 5 of 50 accounts, two of them from builders of competing tools (acc-41, acc-42) and one a vendor-run benchmark; the job is named often but few users report doing it with a listed tool

As data: https://www.anchorterminal.com/api/v1/use-cases/guardrails.gate-tool-calls.json, or the MCP tool get_use_case with id guardrails.gate-tool-calls. Draft, read 2026-11-15 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
