← Back to all sparks
S

Speakeasy

DEVOPS
Velocity10.0

Speakeasy stopped inventorying MCP servers and started adjudicating them.

ai-governanceshadow-mcppolicy-enforcementagent-observabilityidentity-resolution
Current state
Speakeasy ships near-daily platform releases with unusually legible notes — each headline states what changed for a user, not a version number. The current one turns the Shadow MCP page into a single review surface where every server carries an approval state and an automatically gathered evidence dossier: publisher, requested scopes, declared capabilities, maintenance signals, and whether internal teams already talk to it. Decisions enforce on record. Around it, the assistant surfaces have been consolidating: one detail panel for configuration and observation, exact session totals, and canonical identities folding a person's work and personal AI accounts together.
Where it's heading
The arc runs observe, then intercept, now adjudicate. Earlier releases catalogued spend and inventoried shadow MCP servers; the LiteLLM integration moved enforcement to the proxy so a violating prompt dies before inference; this release supplies the judgment layer, doing the research an approver would otherwise do by hand. The supporting work points the same way — prompt-injection scanning of captured skill manifests, risk policies that pause instead of being deleted, identity resolution that reports a whole person rather than an account. Each is a piece a control plane needs before its verdicts can be trusted.
Prediction
Expect approval state to start gating traffic rather than only recording a decision, and the evidence dossier to extend from MCP servers to the skills and assistants already being captured. The rollout flag on the approval workflow suggests general availability is the next step rather than new capability.

Recent moves

  1. 4d ago

    Approve or deny MCP servers with gathered evidence, and pause risk policies without deleting them

    ⚡ SPARK

    This is the point where Speakeasy's shadow MCP work stops being an inventory and becomes a decision surface, completing the observe-intercept-adjudicate arc the last three months traced. Pausing risk policies and canonical identity folding are the operational polish that makes such decisions durable.

  2. 5d ago

    Exact assistant session totals and a hardened dashboard

    Session activity totals are computed over the selected range independent of pagination, so summary numbers stop shifting as pages load — a correctness fix in the analytics that the cost and identity work depends on. Three browser isolation headers close findings from a penetration test.

  3. 6d ago

    Configure and observe assistants from one panel, and see one person behind many accounts

    The assistant detail panel becomes the single place to configure and observe an assistant, with model, concurrency and warm TTL editable in place and triggers showing the traffic they routed. Folding a person's work and personal AI accounts into one identity is the more consequential half: cost analytics that reports a slice of someone's usage understates it.

  4. 6d ago

    Faster assistants, file attachments in chat, and organization names in every language

    Assistants stop sending every MCP tool schema on every call, discovering tools on demand and connecting to servers on first use, which keeps prompt caching viable for large toolsets. A breaking change lands alongside: AWS and GCP KMS update methods no longer accept the fields that defined what a key is.

  5. 8d ago

    Assistants can see images from Slack, and skills are scanned for prompt injection

    Skill manifests are scanned for prompt injection at capture time with findings attached to skill details, extending the skill-capture pipeline into the risk analysis one. Slack-triggered assistants can now read attached images and fetch ones referenced later in a thread.

  6. 10d ago

    Device Agent is out of preview, with a one-step signed macOS installer

    Device Agent leaves preview with a signed, notarized macOS package that installs through normal MDM, which is the difference between a pilot and a fleet rollout. The legacy OAuth proxy is removed in the same release — clients still holding proxy refresh tokens must re-authorize.