LoginSign Up
← Back to all sparks
S

Speakeasy

DEVOPS
Velocity10.0

Speakeasy moved enforcement to the proxy: a bad prompt now dies before inference, not after.

ai-governancemcpproxy-enforcementcost-attributionrbacshadow-it
◆Current state
Speakeasy is an AI control plane that meters agent spend and governs MCP servers, and it crossed 1.0 at the end of July by enforcing role-based access on every plan tier instead of gating it behind a plan. The August releases widen what it can see — ChatGPT and ChatGPT Work usage now counts toward tokens under management alongside Claude Code, Codex, and Cursor — and add a second shadow-MCP posture that allows servers by default and names the ones to block. LiteLLM has become a first-class source, with prompt policies checked at the proxy.
◆Where it's heading
The product is moving from observing AI usage to intercepting it. Early releases catalogued spend and inventoried shadow MCP servers; the newest ones stop a policy-violating prompt before it reaches a model, and turn a bypass approval into a project-wide unblock rather than a private exception. Coverage breadth and enforcement point are being pushed together, which is what a control plane needs in order to be authoritative rather than advisory.
◆Prediction
Expect more first-class sources onboarded through the same proxy-ingestion pattern, and the project assistant to keep gaining the ability to act on findings rather than list them. The entries say nothing about enforcement latency or what happens when the proxy check fails, which is the open question for anything sitting in the request path.

◆Recent moves

  1. 9d ago

    ChatGPT spend shows up in your cost analytics, and shadow MCP policies can allow by default

    ChatGPT and ChatGPT Work usage now counts toward tokens under management and shows up in usage and cost views split by product, closing a visibility gap next to Claude Code, Codex, and Cursor. The allow-by-default shadow MCP posture introduced here is the policy model the following release makes operable.

  2. 9d ago

    Run an allow-by-default shadow MCP policy end to end

    Allow-by-default shadow MCP policies become usable from the surfaces people work in: the inventory shows what a rule blocks, per-server actions become Block or Unblock, and approving a bypass unblocks the server for the whole project. It finishes the posture shipped days earlier, on the premise that a workable default beats a stricter one nobody keeps.

  3. 9d ago

    Integrate and govern every model call that runs through a LiteLLM

    ⚡ SPARK

    Enforcement moves ahead of inference: with LiteLLM as a first-class source, prompt policies are checked at the proxy and a violating prompt is stopped before it reaches a model. This is the point where Speakeasy stops describing AI usage and starts controlling it, which is what the preceding releases were building toward.

  4. 12d ago

    Access controls apply on every plan, and external OAuth setup configures itself

    ⚡ SPARK

    The 1.0 release made role-based access grants unconditional across plan tiers with the rollout flag removed, and taught external OAuth setup to discover and validate provider metadata itself. Making access control a platform guarantee rather than a plan feature is what let the later governance work assume enforcement is always present.

  5. 12d ago

    Ask the project assistant and get an answer cited from the docs, and read billing for whatever date range you pick

    The project assistant can cite the documentation page an answer came from, and the billing usage card now follows whatever date range you select instead of snapping back to the full cycle. Citations matter more than they look here: an assistant that advises on spend and risk has to be checkable.

  6. 12d ago

    Connect Claude Code and VS Code without client registration, and serve an MCP server straight from your domain root

    MCP clients that identify by URL, including Claude Code and VS Code, can finish OAuth without Dynamic Client Registration, and custom domains gained a root route so a server answers at the bare domain. Connection friction removal, which is the groundwork for governing the servers people actually end up connecting.