Anthropic Claude

AttestaToolGate evaluates tool_use content blocks from Claude’s API response through the Attesta approval pipeline before your application executes the requested tool. This gives you full control over which tool calls Claude is allowed to make.

Installation

pip install attesta[anthropic]

API Reference

AttestaToolGate

AttestaToolGate(attesta, risk_overrides=None)

Parameter	Type	Description
`attesta`	`Attesta`	A configured Attesta instance
`risk_overrides`	`dict[str, str] \| None`	Optional mapping of `{tool_name: risk_level}` to force specific risk levels

Methods:

Method	Signature	Returns	Description
`evaluate_tool_use`	`async (tool_use_block) -> tuple[bool, ApprovalResult]`	`(approved, result)`	Evaluates a `tool_use` content block
`make_denial_result`	`(tool_use_id, reason) -> dict`	`tool_result` block	Creates a `tool_result` block with `is_error=True`

evaluate_tool_use accepts both the Anthropic SDK’s ToolUseBlock objects (attribute access) and plain dict representations (key access). This means it works with both the official SDK and raw API responses.

Full Example

import asyncio
from anthropic import Anthropic
from attesta import Attesta
from attesta.integrations.anthropic import AttestaToolGate

# Configure
client = Anthropic()
attesta = Attesta.from_config("attesta.yaml")
gate = AttestaToolGate(
    attesta,
    risk_overrides={"run_bash": "critical"},
)

# Define tools for Claude
tools = [
    {
        "name": "run_bash",
        "description": "Execute a bash command on the server",
        "input_schema": {
            "type": "object",
            "properties": {
                "command": {"type": "string", "description": "The bash command to run"},
            },
            "required": ["command"],
        },
    },
    {
        "name": "read_file",
        "description": "Read the contents of a file",
        "input_schema": {
            "type": "object",
            "properties": {
                "path": {"type": "string", "description": "File path to read"},
            },
            "required": ["path"],
        },
    },
]


async def chat_with_claude():
    messages = [{"role": "user", "content": "Delete all log files older than 30 days"}]

    while True:
        response = client.messages.create(
            model="claude-sonnet-4-20250514",
            max_tokens=1024,
            tools=tools,
            messages=messages,
        )

        # Process each content block
        tool_results = []
        for block in response.content:
            if block.type == "text":
                print(block.text)

            elif block.type == "tool_use":
                # Gate the tool call through Attesta
                approved, result = await gate.evaluate_tool_use(block)

                if approved:
                    # Execute the tool
                    output = execute_tool(block.name, block.input)
                    tool_results.append({
                        "type": "tool_result",
                        "tool_use_id": block.id,
                        "content": output,
                    })
                else:
                    # Send denial back to Claude
                    reason = f"risk: {result.risk_assessment.level.value}"
                    denial = gate.make_denial_result(block.id, reason=reason)
                    tool_results.append(denial)

        if response.stop_reason == "end_turn":
            break

        # Continue the conversation with tool results
        messages.append({"role": "assistant", "content": response.content})
        messages.append({"role": "user", "content": tool_results})


asyncio.run(chat_with_claude())

How tool_use Evaluation Works

When Claude responds with a tool_use block:

{
  "type": "tool_use",
  "id": "toolu_01ABC",
  "name": "run_bash",
  "input": {"command": "find /var/log -mtime +30 -delete"}
}

evaluate_tool_use extracts the tool name and input, builds an ActionContext, and runs the full Attesta evaluation:

# Internally, AttestaToolGate builds:
ctx = ActionContext(
    function_name="run_bash",
    kwargs={"command": "find /var/log -mtime +30 -delete"},
    hints={},               # or risk_override if configured
    agent_id="claude",      # always set to "claude"
)

The agent_id is always set to "claude" on all contexts created by AttestaToolGate. This allows the trust engine to build a trust profile specific to Claude’s tool-calling behavior over time.

Denial Messages

When a tool call is denied, make_denial_result creates a tool_result block that Claude understands:

denial = gate.make_denial_result(
    tool_use_id="toolu_01ABC",
    reason="risk: critical",
)

# Returns:
{
    "type": "tool_result",
    "tool_use_id": "toolu_01ABC",
    "content": "[ATTESTA DENIED] risk: critical. Please suggest an alternative approach or explain why this action is necessary.",
    "is_error": True,
}

The is_error=True flag tells Claude that the tool call failed. The denial message asks Claude to suggest alternatives, which typically produces more helpful responses than a silent failure.

Agentic Loop with Gating

Here is a complete agentic loop that processes multiple tool calls per turn and handles both approvals and denials:

import asyncio
from anthropic import Anthropic
from attesta import Attesta
from attesta.integrations.anthropic import AttestaToolGate


async def agentic_loop(user_message: str):
    client = Anthropic()
    attesta = Attesta.from_config("attesta.yaml")
    gate = AttestaToolGate(attesta)

    messages = [{"role": "user", "content": user_message}]

    for turn in range(10):  # Max 10 turns
        response = client.messages.create(
            model="claude-sonnet-4-20250514",
            max_tokens=4096,
            tools=tools,
            messages=messages,
        )

        if response.stop_reason == "end_turn":
            # Claude is done -- extract final text
            for block in response.content:
                if block.type == "text":
                    return block.text
            break

        # Process tool_use blocks
        results = []
        for block in response.content:
            if block.type != "tool_use":
                continue

            approved, eval_result = await gate.evaluate_tool_use(block)
            if approved:
                output = await execute_tool_async(block.name, block.input)
                results.append({
                    "type": "tool_result",
                    "tool_use_id": block.id,
                    "content": str(output),
                })
            else:
                level = eval_result.risk_assessment.level.value
                results.append(
                    gate.make_denial_result(block.id, reason=f"risk: {level}")
                )

        messages.append({"role": "assistant", "content": response.content})
        messages.append({"role": "user", "content": results})

    return "Max turns reached"

Risk Overrides

Force specific risk levels for known-dangerous tools:

gate = AttestaToolGate(
    attesta,
    risk_overrides={
        "run_bash": "critical",     # Shell commands always critical
        "write_file": "high",       # File writes always high
        "read_file": "low",         # File reads always low
    },
)

When a tool name matches a risk override, the hint {"risk_override": "critical"} is passed to the scorer, which forces the specified risk level regardless of the heuristic score.

CrewAI

Gate CrewAI task outputs

OpenAI Agents SDK

Approval handlers and guardrails

Code Integrations

No-Code Platforms

Anthropic Claude

Installation

API Reference

AttestaToolGate

Full Example

How tool_use Evaluation Works

Denial Messages

Agentic Loop with Gating

Risk Overrides

CrewAI

OpenAI Agents SDK

​Installation

​API Reference

​AttestaToolGate

​Full Example

​How tool_use Evaluation Works

​Denial Messages

​Agentic Loop with Gating

​Risk Overrides

CrewAI

OpenAI Agents SDK

Installation

API Reference

AttestaToolGate

Full Example

How tool_use Evaluation Works

Denial Messages

Agentic Loop with Gating

Risk Overrides