# Computer use (computer_toolset_20260801, computer_20251124, computer_20250124)

Status: DOCUMENTED · LIVE_VERIFIED 2026-09-18 for request/response shapes (h1/h2 computer_20251124 on claude-sonnet-4-6; h3 computer_20250124 on haiku 4.5; h4 computer_toolset_20260801 on claude-sonnet-5; h5/k17/k17b/k22 error cases). No real display was driven — the loop below returns a 1×1 PNG placeholder. Sources: Computer use tool · Tool reference — client toolsets · Beta Messages reference · Release notes 2026-08-19. Last verified: 2026-09-18.

# Versions

Platforms: toolset on Claude API + Google Cloud; AWS/Bedrock/Foundry only the beta versions. Not available in Managed Agents.

# Actions / members

Coordinates are in the pixel space of the screenshot you returned (toolset) or of display_width_px × display_height_px (2025 versions). If you downscale screenshots, scale coordinates back up; macOS Retina = ×2. Image limits: Opus 4.7+ models 2576 px long edge / 4784 visual tokens; earlier 1568 px / ~1.15 MP; the toolset rejects oversized images instead of downscaling. > 20 images per request → stricter per-image limits.

Live shapes:

// h1 (computer_20251124, sonnet 4.6)
{"type":"tool_use","id":"toolu_0185…","name":"computer","input":{"action":"screenshot"},"caller":{"type":"direct"}}
// h4 (computer_toolset_20260801, sonnet 5) — preceded by an adaptive `thinking` block
{"type":"tool_use","id":"toolu_0181…","name":"screenshot","toolset_name":"computer","input":{},"caller":{"type":"direct"}}

# Batch actions (toolset) and results

Several member tool_use blocks per turn = batch. Run in order, stop at the first failure, answer every block: success → normal result; failed → is_error:true + description; later ones → is_error:true + exactly Not executed: an earlier computer action in this turn failed. Every member result must echo "toolset_name": "computer" and may contain only text/image blocks. cache_control on any tool_use/tool_result of a batch acts once at the end of the batch. Limit to one action per turn with tool_choice: {type: auto, disable_parallel_tool_use: true}.

Rejected on the toolset entry (400): name (k22: "name is not accepted on a toolset entry…"), display_*, enable_zoom, strict, input_examples, defer_loading on the entry, tool_choice type tool, legacy fine-grained header, a second entry / another tool named computer, configs disabling every member.

# Client-side loop (Python) — examples/anthropic/computer-use/computer_use_loop.py

def sampling_loop(prompt, max_iterations=10):
    messages = [{"role": "user", "content": prompt}]
    for _ in range(max_iterations):
        resp = client.beta.messages.create(model=MODEL, max_tokens=4096, tools=TOOLS, messages=messages,
                                           betas=["computer-use-2025-11-24"])      # no betas for computer_toolset_20260801
        messages.append({"role": "assistant", "content": resp.content})
        results, failed = [], False
        for b in resp.content:
            if b.type != "tool_use" or not (b.name == "computer" or getattr(b, "toolset_name", None) == "computer"):
                continue                                    # dispatch bash / text editor / custom tools here
            action = b.name if getattr(b, "toolset_name", None) else b.input["action"]
            r = {"type": "tool_result", "tool_use_id": b.id}
            if getattr(b, "toolset_name", None): r["toolset_name"] = "computer"
            if failed: r.update(content=NOT_EXECUTED, is_error=True)
            else:
                try: r["content"] = execute_action(action, b.input)   # screenshot -> [{"type":"image",...}], else "OK"
                except Exception as e: r.update(content=f"Error: {e}", is_error=True); failed = True
            results.append(r)
        if not results or resp.stop_reason != "tool_use": return messages
        messages.append({"role": "user", "content": results})

Ran live (2 requests each): step 1 → screenshot request; step 2 with a 1×1 PNG → Claude described a dark screen (h2, stopped at max_tokens 50). execute_action is a stub — connect it to Xvfb/xdotool in a sandboxed VM.

# Security, limits, pricing

• Run in a dedicated VM/container, no credentials, domain allowlist, human confirmation for consequential actions; Anthropic's prompt-injection classifiers scan screenshots (opt-out via support). Screenshots/keystrokes never stored by Anthropic → ZDR-eligible (client tool).
• Toolset definition ≈ 4,500 input tokens (h4: 4,189 with zoom disabled; sonnet 5); 2025 versions: 466–499 system-prompt tokens + ≈735 per definition (h1: 1,843 total; j8 count_tokens haiku: 1,826). Screenshots ≈ 1,000–1,800 tokens each. Prune screenshots in batches; on Fable 5.1 prefer server-side tool-result clearing (client pruning invalidates thinking blocks).
• Thinking: models using the toolset omit thinking text by default — set thinking.display: "summarized" to inspect reasoning. For computer_20251124, docs suggest effort high (Opus 4.7) / medium (Sonnet 4.6, Opus 4.6).

Examples: computer_toolset_20260801.sh, computer_20251124.sh, computer_20250124_haiku.sh, computer_use_loop.py|.ts in examples/anthropic/computer-use/. Test: test_computer_use_beta_screenshot (expensive).