Amazon Workspaces Agent Access

Amazon WorkSpaces Applications — Agent Access

Domain expertise for connecting AI agents to remote Windows desktops on Amazon WorkSpaces Applications (AppStream 2.0) via the managed Agent Access MCP server, and for driving those desktops reliably.

How it works: Agent Access is MCP-only — there is no AWS CLI/SDK command that calls the desktop tools. Agents connect to https://agentaccess-mcp.{region}.api.aws/mcp over Streamable HTTP, SigV4-signed with service name agentaccess-mcp, and call MCP tools (screenshot, left_click, type_text, ...) to drive the desktop. The AWS CLI/SDK is used only for setup — appstream create-streaming-url, fleet/stack configuration. mcp-proxy-for-aws handles the SigV4 signing.

Recommended setup: use mcp-proxy-for-aws (Python) as the transport; it signs each request and manages the DELETE lifecycle. Any MCP client that supports Streamable HTTP + SigV4 works. When running the AWS CLI/SDK setup steps (create-streaming-url, stack/fleet configuration), the AWS MCP server is recommended for sandboxed execution and audit logging.

Guardrail — where this skill's own files live (MCP vs local install)

This skill can be loaded two ways, and they resolve the skill's own bundled files from different places. Determine how the skill was loaded before reading a reference:

• Loaded through the AWS MCP retrieve_skill tool: The skill is not installed on the local filesystem. You MUST fetch each reference via retrieve_skill with the file parameter (e.g. file="references/connection-setup.md"), and use the returned content. Do NOT file_read these paths locally — they do not exist on disk.
• Installed locally (e.g. .kiro/skills/amazon-workspaces-agent-access/ or ~/.claude/skills/amazon-workspaces-agent-access/): Read files from the local skill directory using relative paths.

This distinction applies only to the skill's own packaged files. User data and session artifacts are always read from and written to the user's working directory. Never fetch or write customer data through retrieve_skill.

Key facts agents get wrong (load the reference before answering in detail)

These are HTTP headers / metadata on the MCP connection — not tool parameters, and there is no connect_to_desktop tool. Do not invent tools or parameters; the desktop tools are exactly those in tools-reference.md.

• Connect mode. Selected by the X-Amzn-AgentAccess-Connect-Mode HTTP header (value BLOCKING, the default, or POLLING) — sent on the MCP request alongside the streaming-URL/SAML auth. It is NOT a JSON tool argument.

  • ❌ WRONG (common hallucination): calling a connect_to_desktop tool with a connection_mode: "POLLING" parameter, or a session_id/application_id/user_id argument. None of those exist.

  • ✅ RIGHT: set the X-Amzn-AgentAccess-Connect-Mode: POLLING header. Then tools/list initially returns only the connection_status tool; the agent calls connection_status repeatedly until its returned state is CONNECTED, and only then does tools/list return the full desktop tool set (screenshot, left_click, ...). (details: connection-modes.md)

    python

    # Correct POLLING usage — the mode is an HTTP header on the MCP connection:async with aws_iam_streamablehttp_client(    endpoint="https://agentaccess-mcp.us-east-1.api.aws/mcp",  # use YOUR fleet's region    aws_service="agentaccess-mcp", aws_region="us-east-1",  # region must match the fleet (else 400)    headers={        "X-Amzn-AgentAccess-Streaming-Session-Url": streaming_url,        "X-Amzn-AgentAccess-Connect-Mode": "POLLING",   # header, not a tool arg    },) as (read, write, _):    async with ClientSession(read, write) as session:        await session.initialize()        # tools/list now returns ONLY connection_status until the desktop is up:        while json.loads((await session.call_tool("connection_status", {})).content[0].text)["state"] != "CONNECTED":            await asyncio.sleep(2)        tools = await session.list_tools()   # now the full desktop tool set

• Streaming session (non-domain-joined) is the X-Amzn-AgentAccess-Streaming-Session-Url header. Domain-joined fleets instead pass the SAML assertion + stack ARN via MCP _meta keys aws.agentaccess/workspacesApplicationsSamlAssertion and aws.agentaccess/workspacesApplicationsStackArn. (details: connection-setup.md)

• Expire-on-delete is the X-Amzn-AgentAccess-Expire-Streaming-Session-On-Delete header (true/false; default false). Expiry happens on the client's explicit HTTP DELETE — mcp-proxy-for-aws sends it automatically on clean close. (details: session-lifecycle.md)

• Forwarded tools are namespaced by server: forwarded___<server-name>___<tool-name> (e.g. forwarded___filesystem___read_file) — not forwarded___<tool-name>. (details: tool-forwarding.md)

• COMPUTER_INPUT requires COMPUTER_VISION to also be ENABLED in the stack's AgentAccessConfig. (details: enabling-agent-access.md)

Routing

Security Considerations

• The agent acts under the caller's AWS identity. Every MCP request is SigV4-signed with service agentaccess-mcp; the desktop session runs with those credentials. Grant only the specific agentaccess-mcp actions the agent calls (e.g. InvokeMcp, GetScreenshot, LeftClick, TypeText) and scope them with the agentaccess-mcp:StackArn condition key — avoid a blanket agentaccess-mcp:* or Resource: *. Prefer IAM roles over long-lived users. (Full action list + example: connection-setup.md → IAM permissions.)
• Screenshots can capture sensitive data. COMPUTER_VISION captures whatever is on the desktop — treat screenshots as potentially containing PII or secrets. If screenshot storage is enabled, the S3 bucket must enforce encryption at rest and in transit and least-privilege access: grant the AppStream service principal only what it needs and the connecting agent only s3:PutObject (see enabling-agent-access.md).
• Enable only the capabilities you need. COMPUTER_INPUT, COMPUTER_VISION, and FORWARD_MCP_TOOLS are independent — do not enable input/forwarding on stacks that only need vision.
• Tool forwarding executes code on the fleet. Forwarded MCP servers run on the instance under the session context. Install only trusted servers system-wide, gate with FORWARD_MCP_TOOLS, and scope the CallForwardedTool IAM action by agentaccess-mcp:StackArn (see tool-forwarding.md).
• Keep a human in the loop where warranted. UserControlMode: VIEW_STOP lets an observer watch the live session and stop the agent. Treat agent-driven desktop actions as capable of arbitrary UI operations.
• Audit with CloudTrail. Agent session events are logged; tool calls are CloudTrail data events and require a trail configured to log them. Create a trail with agentaccess-mcp data events enabled, encrypt it with SSE-KMS, and add CloudWatch alarms for anomalous patterns (e.g. high screenshot volume, unexpected TypeText, repeated auth failures). If screenshot storage is enabled, turn on S3 server access logging for the bucket.
• Protect federation material. For domain-joined (SAML) fleets, safeguard the IdP signing certificate and the IAM SAML provider/role trust policy, and do not log the SAML assertion. Traffic is HTTPS + SigV4 — never disable TLS verification.
• Treat typed input as potentially sensitive. type_text can enter secrets (passwords, tokens); these may then appear in screenshots, screenshot-storage S3, and CloudTrail data events. Avoid typing long-lived secrets into the desktop where possible, and restrict who can read those sinks.
• Refer to the current Agent Access documentation and AWS security best practices for the latest guidance.

Note: Regional endpoints, feature availability, and quotas change. When precision matters, confirm against the current Agent Access MCP server documentation. The references focus on the values and gotchas that are easy to get wrong.