LoginSign Up
← Back to all sparks
S

Speakeasy

DEVOPS
Velocity10.0

Speakeasy makes AI agents first-class principals: server identity modes, Slack as a work identity, and directory group role mapping

mcpai-agentsidentityenterprisesecuritygovernance
◆Current state
Speakeasy is an enterprise AI agent governance platform shipping at a high cadence — multiple minor versions per week. The current window covers two structural areas: MCP identity management (remote servers now declare User/Agent/No Identity modes, agents own their upstream accounts, wildcard workload admission for agent fleets) and enterprise directory integration (Slack as a work identity, directory group-to-role mapping). The risk policy engine gained scoped targeting for specific MCP servers and tools in the same release window.
◆Where it's heading
Speakeasy is building the control plane for enterprise AI agent governance: making agents governable by the same tools that manage human employees — identity, directory sync, role assignment, risk policies, and audit trails. The workload assertion grant (a short-lived MCP session from a platform-issued identity token) is the primitive that makes agent fleets manageable at scale. Each release tightens the coupling between AI agent identity and enterprise identity providers (Slack, OIDC, directory groups).
◆Prediction
The next material release will likely extend the Shadow AI scan's person-level visibility into policy actions (block/allow per person per tool), or introduce agent groups as a first-class concept analogous to directory groups — the wildcard workload admission is a precursor to that. Expect tighter Okta Cross App Access integration as that feature exits the readiness/preview stage.

◆Recent moves

  1. 1mo ago

    The latest models reach the pickers, and published plugins work on OpenCode 2 again

    Claude Opus 5.5, Claude Fable 5.1, GPT-6 Astra/Sol/Luna, Gemini 3.8 Flash, Grok 4.7, DeepSeek V4.1 Flash, and several others are added to all model pickers. OpenCode 2 plugin compatibility is restored — published plugins were silently failing to register MCP servers and skills on OpenCode 2. The command palette now ranks results by intent rather than substring match, using the organization's existing model key.

  2. 1mo ago

    Risk policies can target chosen MCP servers, gateways, and tools instead of everything

    Risk policies can now target specific MCP servers and their tools rather than applying everywhere. The scope step offers a two-pane picker with per-server tool rules and an Inspect matrix, and organization admins can disable the Shadow AI scan on specific device agent fleets. This follows the identity management work in v1.31.1 — granular risk policies require knowing which server is which, which now has a clean answer.

  3. 1mo ago

    Directory groups grant roles on their own, and Slack accounts resolve to the people behind them

    ⚡ SPARK

    Slack becomes a work identity in Speakeasy: admins connect workspaces, sync directories, and map Slack members to existing personnel so an account posting in Slack and a person in the platform are one record. Directory group-to-role mapping arrives in the same release: matching members receive roles automatically alongside their direct assignments, without manual list maintenance.

  4. 1mo ago

    Remote MCP servers choose who they act as, and a workload fleet can be admitted by one wildcard

    ⚡ SPARK

    Remote MCP servers now declare whose identity they act under — User, Agent, or No Identity — configured from one authentication panel with managed credentials. Workload fleet admission gains wildcard support: a single entry admits every subject beneath an issuer-permitted path and resolves the agent they inherit policy from. Two new RBAC scopes gate who can manage the workload trust policy at all.

  5. 1mo ago

    A detected AI tool now opens onto the people running it

    The Shadow AI inventory now pivots from tool to person: opening a detected AI tool expands it into the enrolled users running it, with devices, versions, signals, and first and last sightings. This inverts the previous identity-page view (start from a person, see their tools) — an access decision usually starts from the tool, and now so can the workflow.

  6. 1mo ago

    Agent key creation survives agents with many servers, and delegated credentials recheck authority before they refresh

    Two reliability fixes for agent key management: creation no longer times out on agents fronting many MCP servers (permission discovery collapses to one request), and delegated credentials now recheck live authorization before refreshing rather than trusting a cached picture of who authorized them. The latter is a correctness fix for the delegated identity model introduced in earlier releases.