Anthropic¶

Available · Messages API tools · no SDK needed

The renderers return plain dicts. Import the functions by name (from agent_adapter.integrations.anthropic import ...), so the anthropic SDK name stays free.

Functions¶

Function (`agent_adapter.integrations.anthropic`)	Returns
`anthropic_tools(registry)`	every tool in Messages API format
`anthropic_tool(spec)`	one tool: `{"name", "description", "input_schema"}`
`tool_result_blocks(results)`	one `tool_result` block per `ToolResult`, in order
`INTEGRATION`	the integration object: `export_tools(registry)`, and `bind_context(block)` puts the `tool_use` id into `ctx.call_id`

Round trip¶

Pass the rendered list as tools=. Run the tool_use blocks the model sends through one Executor, never with spec.invoke: invoke skips approval, budgets, run guards, the cache and the ledger, so a confirm=True tool would run unapproved.

from agent_adapter import Executor, ToolCall
from agent_adapter.integrations.anthropic import anthropic_tools, tool_result_blocks

executor = Executor(registry)  # once per process; no approval port, so confirm tools are denied
tools = anthropic_tools(registry)  # pass as tools= to messages.create
calls = [
    ToolCall(name=block.name, arguments=block.input, id=block.id)  # input is a dict
    for block in message.content
    if block.type == "tool_use"
]
results = await executor.run_many(calls)  # one ToolResult per call, in order
reply = {"role": "user", "content": tool_result_blocks(results)}

sequenceDiagram
  participant App
  participant A as Anthropic API
  participant Ex as Executor
  App->>A: messages.create(..., tools=anthropic_tools(registry))
  A-->>App: content blocks of type "tool_use": id, name, input (dict)
  App->>Ex: run_many([ToolCall(name, input, id), ...])
  Ex-->>App: ToolResults in order (approval, budgets and guards applied)
  App->>A: user message of tool_result_blocks(results)

Step	Notes
Executor	build it once, with `approval=` if any tool has `confirm=True`; with no port such a call is denied (approval)
One call	`await executor.run_call(ToolCall(...))` gives one `ToolResult`
Whole loop	`run_tool_loop` with `AnthropicModel` sends the calls and answers them for you

Answering tool calls¶

tool_result_blocks turns results into the next user message. An ok value is sent as text (strings as they are, anything else as JSON); an error sets is_error and sends {"error": {"kind", "message"}}; a pending approval sends approval required: <id>. examples/anthropic_results.py answers two calls, one of which fails.

What the schema looks like¶

input_schema is the plain schema: $ref inlined, title dropped, const turned into a one-value enum. Optional fields stay optional and keep their default.