QWED

Type to search pages, engines and advisories.

    Looking for API details?Search the docs

    Solutions·AI agents

    Runtime verification for agent actions, before they run.

    Agents call tools, talk to other agents and spend money. QWED adds a check at each of those points: tool calls before they run, messages between agents, and checkout totals before payment.

    What QWED checks here

    • Tool calls
    • Agent outputs
    • Messages between agents
    • Checkout and payment

    With Open Responses, QWED A2A, QWED UCP and MCP server

    Agents act at machine speed. So do their mistakes.

    An agent that is wrong does not pause to ask. It acts.

    Without QWEDNobody checked any of itWith QWEDEach one checked before it is used
    A destructive tool callAn agent calls a shell tool with rm -rf /.Open ResponsesBLOCKED · DENYA blocked tool pattern: denied, so your agent does not run the call.
    A total that does not add upOne agent tells another that items worth 150.00 come to 999.99.QWED A2AVERIFIED · DENYThe items come to 150.00, so the message is held.
    Tax applied wrong at checkoutTax of 8.25% on $100 is charged as $8.00 just before payment goes through.QWED UCPVERIFIED · DENY8.25% of $100 is $8.25, so payment is held.

    What is proven correct is admitted, with a record of why. QWED returns the decision; your integration acts on it.

    What QWED checks

    Four packages cover the places where an agent acts. None of them asks a model to judge the result.

    • Open Responses

      Tool calls

      ToolGuard blocks dangerous tools and patterns, ArgumentGuard checks argument types and ranges, and SafetyGuard flags personal data, prompt injection and budget.

    • Open Responses

      Agent outputs

      MathGuard recomputes totals, SchemaGuard checks JSON against a schema, and StateGuard checks that each state change is allowed.

    • A2A

      Messages between agents

      A trust boundary with allowlists and rate limits, exact checks on money totals, and a signed attestation on every decided result.

    • UCP

      Checkout and payment

      Total must equal subtotal minus discount plus fulfilment, tax and fees, and the checkout must move through valid states.

    • MCP server

      Exact answers in MCP clients

      Lets Claude Desktop, VS Code and other MCP clients run Python calculations in a restricted subprocess instead of guessing.

    Where it fits

    1. Your AI

      Decides to call a tool, send a message or complete a purchase.

    2. QWED

      Checks the call, the message or the totals against your rules.

    3. Your systems

      Your integration passes only admitted actions to your tools, other agents and payment systems.

    Good to know

    • Open Responses blocks tools from a list you configure. It is not a sandbox, and its pattern scanning is a heuristic.
    • In QWED A2A, only finance checks can return verified; logic and code checks alone are not enough to act on.
    • The MCP server runs code with the server’s privileges, so connect only clients you trust.

    Get started

    • Open Responses

      v0.6.1Beta

      Guards for agent tool calls and outputs. Works with the OpenAI Responses API, LangChain and LlamaIndex.

      pip install qwed-open-responses
    • QWED A2A

      v0.3.0Alpha

      A zero-trust interceptor for messages between agents, using the Agent-to-Agent protocol.

      pip install qwed-a2a
    • QWED UCP

      v0.3.0Alpha

      Checks for Universal Commerce Protocol checkouts before the money moves.

      pip install qwed-ucp
    • MCP server

      v0.2.3Beta

      Verification tools for Claude Desktop, VS Code and other MCP clients.

      pip install qwed-mcp

    Building with AI agents?

    Tell us what your AI does and where a mistake would hurt.

    Contact

    Guides for this field

    Other fields