August 2026 Summaries

Speakeasy has released its Platform MCP, a first-party Model Context Protocol server that lets organization administrators govern and manage MCP servers, projects, plugins, skills, access controls, OAuth readiness, diagnostics, observability, and documentation from agent environments such as Claude, Cursor, Codex, and OpenCode. The initial release focuses on the MCP gateway, enabling users to search and register catalog or remote servers, manage project configurations, assess connection status, distribute servers and skills through plugins, and investigate failures across Speakeasy, providers, and clients. Access is currently limited to eligible organizations with an active org-admin grant and requires activation through a Speakeasy contact and installation via dashboard plugins or native configuration. Certain sensitive and destructive actions, including OAuth consent and secret setup, identity-provider attachment, plugin creation or deletion, organizational changes, and server authoring, remain restricted to human-controlled dashboard workflows.

Speakeasy has introduced session portability for AI Control Plane customers, allowing engineers to transfer work between Codex, Claude Code, and Cursor when changing tools, licenses, or encountering token limits. Running through the macOS Speakeasy Device Agent in public beta, the feature reads existing local agent transcripts, creates a handoff containing the task, recent conversation turns, touched files, and latest assistant response, then launches the chosen tool in the same project directory. Users can move sessions through a menu-bar “Continue session” workflow or CLI commands, while target tools receive the context as a new prompt or, for Cursor, through a prefilled SESSION_HANDOFF.md file. Speakeasy states that session content remains on the device and is never uploaded; its control plane records only content-free audit events identifying the user, source and destination tools, and device. The release does not recreate original chats or import Cursor’s native history, and excludes features such as share links, dashboard continuation, and native session-file resume.

Speakeasy has introduced MCP Approval Flow within its AI Control Plane to help organizations turn blocked connections to unreviewed Model Context Protocol servers into documented security decisions. When a shadow MCP policy blocks an agent’s tool call, the user can submit a justification through a link without needing dashboard access, while designated reviewers receive a server profile containing automated evidence about its publisher, authentication requirements, capabilities, maintenance status, organizational usage, and prior reviews. Security teams can approve or deny requests and limit approvals to an individual, team, or project, with each decision and rationale recorded directly in the blocking policy. Optional research agents can collect cited information from independent sources but cannot make approval decisions. After approval, daily checks compare relevant server permissions and advisories against the original review snapshot, flagging changes for renewed evaluation while emphasizing that narrow access grants remain important because interface-based comparisons cannot detect all behavioral changes.

Speakeasy’s comparison of its AI Control Plane with Databricks Unity AI Gateway argues that the products address enterprise AI governance from different starting points: Databricks focuses on routing and governing model traffic and registered MCP services within the Unity Catalog and lakehouse ecosystem, while Speakeasy focuses on identity-based control of tool use, agent behavior, and AI clients across an organization. Unity AI Gateway, generally available in August 2026, provides centralized model access, token-level cost attribution, budgets, rate limits, traffic routing, catalog-based permissions, and audit logging, though several features including Smart Routing, service policies, and agent services remain in beta. Speakeasy uses MCP gateways, endpoint-deployed agent hooks, and directory identity integrations to inspect prompts, tool calls, shell commands, retrievals, and unregistered MCP servers, emphasizing company-wide discovery, policy enforcement, and visibility into personal or off-platform AI usage. The comparison concludes that Unity AI Gateway is better suited to organizations with Databricks-centered AI estates and lakehouse workloads, whereas Speakeasy is positioned for security and IT teams seeking governance across SaaS tools, internal APIs, developer clients, and other systems outside a single data platform; the products can also be used together because they govern different traffic paths.

Enterprise AI governance requires a control layer that can connect user identity and permissions with every model call, tool call, real-time policy decision, and correlated audit record across an organization’s AI environment. Hyperscalers such as AWS, Microsoft, and Google offer governance features within their own cloud platforms but lose visibility when organizations use multiple model providers or external tools. Enterprise platforms and GRC vendors, including ServiceNow and Salesforce, can define policies and manage risk but generally operate above the traffic layer, making real-time enforcement difficult without building new infrastructure. LLM gateways and MCP security tools can enforce controls within their specialized layers but often lack broader identity visibility and cross-layer correlation. The analysis argues that these limitations are structural because incumbent vendors each control only part of the AI stack, while comprehensive governance requires an AI control plane positioned across the connections between those layers.

Speakeasy announced full support for GitHub Copilot in VS Code on its AI Control Plane, extending the identity, security, governance, and observability capabilities already available for Claude Code, Cursor, Codex, and OpenCode. Organizations can apply existing policies to Copilot tool and MCP use, including permissions, blocking, and auditing, while consolidating session activity and agent investigations into a single dashboard rather than separate vendor consoles. The integration instruments developers’ current Copilot installations without requiring a new extension, migration, or workflow changes. Customers can onboard existing Copilot licenses through Speakeasy’s established coding-agent setup process, although the release does not cover Microsoft 365 Copilot or Agent 365.

Speakeasy has introduced opt-in automatic session refresh for remote MCP sessions running through its AI Control Plane MCP gateway, aiming to prevent agents from failing when access or refresh tokens expire during periods of inactivity. When enabled during service connection, the gateway renews expiring access tokens and uses refresh grants daily to help keep integrations such as Slack, GitHub, and Linear active within providers’ inactivity limits. Users can instead refresh connections manually and view known token expiry information, while providers that do not report expiry are treated as having unknown expiration rather than receiving estimated deadlines. The feature operates on a best-effort basis, so revoked credentials or provider policies may still require reconnection, and organization administrators can set refresh policy as disabled, user-controlled, or required.

Speakeasy has introduced a LiteLLM integration that uses its AI Control Plane to enforce organizational policies on model requests routed through LiteLLM’s proxy to providers such as OpenAI and Anthropic. Through LiteLLM’s Generic Guardrail API, prompts are checked before inference, allowing policy-violating content such as leaked credentials to be blocked before reaching a provider or generating costs, while responses are recorded for later risk analysis. The integration sends metadata-only OpenTelemetry data, including model, token, cost, and duration information, to Speakeasy’s Observe and Costs features without transmitting prompt content through telemetry. It also maps virtual-key email addresses to organization members to attribute enforcement actions, findings, and spending to individuals, supports configurable fail-open or fail-closed behavior during platform outages, and requires only environment variables and a guardrail configuration rather than custom callback code.

Speakeasy has joined the Linux Foundation’s Agentic AI Foundation, which governs open standards and projects for agentic AI including the Model Context Protocol, the goose agent framework, and AGENTS.md. The company says it joined to help advance MCP’s enterprise readiness by contributing production experience from its MCP gateway and its large-scale generation of MCP servers from OpenAPI contracts. Its planned contributions focus on identity and authorization, compliance-oriented auditing, human approvals, efficient tool design, security defenses against agent-specific risks such as prompt injection and tool poisoning, and interoperability among AAIF projects. Speakeasy states that membership will not immediately change its products, but will give customers a path for enterprise requirements to influence open protocol development rather than relying on proprietary solutions.

Speakeasy has added OpenCode, Anomaly’s open-source, provider-agnostic AI agent harness, as a fully supported agent on its AI control plane alongside Claude Code, Codex, and Cursor. The integration enables organizations to deploy OpenCode with centrally managed configurations for approved MCP servers, models, and settings, while monitoring identities, model usage, token counts, and costs in a unified dashboard. It also applies existing enterprise risk policies to OpenCode tool calls, including prompt-injection detection, destructive-command blocking, and custom rules such as restrictions on force pushes or production credentials, with sensitive credentials and MCP URLs redacted before storage. Organizations can begin by adding OpenCode through Speakeasy’s dashboard, selecting an enforcement level, and reviewing existing policies that automatically apply to the agent.

Speakeasy is adopting the newly released Agent Plugins 1.0 standard as the portable foundation for distributing AI agent skills and MCP server configurations across different clients such as Cursor, Codex, and future compatible tools. The company argues that fragmented, client-specific packaging creates costly duplication, version drift, and limited tool choice for enterprise teams, while the standard defines a minimal common package containing a manifest, skills, and MCP definitions without attempting to standardize marketplaces, installation, permissions, or credentials. Speakeasy plans to compete in the layers above the package format, including registries, policy-based distribution, versioning, governance, compliance, and usage analytics, while allowing customers to export open-standard packages without vendor lock-in. It emphasizes that credentials are excluded from packages and handled through client-side OAuth, and that curated plugins remain valuable in enterprises because they provide role-specific, approved tools and access rather than relying on ungoverned discovery. Because the specification is new and incomplete, Speakeasy is pinning support to a specific revision, retaining provider-native functionality for unsupported features such as hooks or certain private credential models, and including legacy manifests for older clients.

Speakeasy has added support for the Agent Plugins 1.0 standard, an open format backed by organizations including OpenAI, AWS, Microsoft, GitHub, Cursor, and Vercel that combines agent skills and MCP server definitions into a portable plugin package. Compatible Speakeasy plugins now publish once for use across supporting agents such as Cursor and Codex while preserving existing marketplace identities and legacy installation manifests for older versions. The standard uses a minimal manifest, standardized skills and MCP directories, and agent-specific extensions where needed, while leaving distribution, permissions, and credential management to individual agents. Speakeasy validates packages against a fixed version of the specification and prevents partial exports when a plugin contains unsupported configurations. Portable packages exclude credentials entirely, relying on agent-managed OAuth; plugins requiring embedded secrets, custom credential headers, or non-OAuth private MCP servers remain on provider-native packaging. Claude Code, OpenCode, and observability hooks continue to use native formats for now, while users can automatically receive the new format on their next compatible publish or download a standards-compliant ZIP for other supported agents.