Roadmap
The work in progress across the Speakeasy AI control plane, mirrored from our public roadmap. Quarters are what we are planning toward, not commitments. Vote on anything here and we read it.
6
In progress
Being built now, across every pillar.
7
Planned
Scoped and scheduled for a quarter.
53
Shipped
Live in the product, with release notes.
01 / Connect
MCP servers, SaaS connectors, and internal APIs provisioned through one catalog and reached through one gateway URL, with the clients your teams already run.
8 items
Device Agent shows which plugin MCP servers need sign-in and connects them with one click, without changing any underlying OAuth flows.
Host a monolithic gateway endpoint that exposes whatever set of MCPs the authorized user has access to
Customer-defined sensors and signals that classify agent conversations into time-series insights, with rich dashboard analytics, headless configuration, and exportable readings.
Platform MCP capabilities so a compliance team's agent can implement ISO 27001 controls: policy redline + matching AICP configuration from one governed conversation.
Allow customers to keep their secrets in 1Password with Speakeasy only utilizing them through refs
Investigate and (if feasible) add Antigravity to supported MCP clients.
Enable individual AI accounts with no-training evidence and onboarding enforced through supported hooks.
Add Zed as a supported MCP client surface for MoonPay POC follow-up.
02 / Control
Who can use what, under which conditions. Scoped access by team, role, and tool, with the identity provider you already run as the source of truth.
1 item
Allow admins to cut off access to AI by user/session/agent
04 / Observe
Sessions, tokens, tool calls, and risk findings by team, client, and user, exported to the stack your security and finance teams already run.
4 items
Add a real compliance/attestation POST (agent.reportState) so device coverage moves from "we saw a heartbeat" to "policy X is verifiably applied on this device," upgrading MDM coverage buckets and Drata/Vanta evidence.
/mcp-sessionsRedesign the `/mcp-sessions`: connection trends, failure hot spots, kill switches, client restrictions and better search.
Track and visualize the lineage of agent runs
MCP elicitation · URL modeShipped
Roadmap items that are in the product today. The changelog has the release-by-release detail.
53 items
Anthropic Compliance APIPreviously known as Anthropic Compliance API integration.
Graduate Assistants from Early Access to Beta
Role plugins · day oneAutomatic role-based MCP server delivery, with client activation capabilities under investigation.
Let customers bring their own Openrouter keys for the Assistants, Risk Policies and other inference capabilities on the platform.
Turn shadow-MCP blocks into a desktop notification with a one-click access request that lands in the admin Approval Requests queue.
if an MCP server is hosted or proxied by speakeasy, when authenticating, it's possible for a user to authenticate specific annotations of tools, and/or specific tools. Does not override permissioned based access to tools through RBAC.
Generate the events and metrics for our new set of tables, and introduce all the catalogs necessary to query the new tables via the Semantic API.
The query builder over the semantic catalog — the surface people use to ask questions of a dataset, and where queries are saved.
Better documentation, homebrew support, ability to check deployment status and toml/yaml config support.
Custom docs · custom logosAllow users to specify external docs and custom logos for public MCP install pages
Process all customer deployments under 10 seconds
Gram FunctionsSupport agentic tools defined in code in Python and Typescript
Improving the search and table views in /logs/mcp
Ingress rules · tailnet-awareAdd network-level access controls for MCP servers via configurable ingress rules and tailnet-aware access.
Enhance Hosted MCP Server endpoints to support filtering of tools and resources by a tool parameter.
Hooks coverageExtend AI Control Plane hooks coverage to OpenCode for MoonPay POC follow-up.
Offer visibility into personal account usage on company devices
First-party Platform MCP plus published agent plugins (MCP and skills) that let customer org admins run day-to-day Gram administration headlessly from Claude, Cursor, Codex and other harnesses.
Tunnelling functionality in order to make customers VPC/On prem services data accessible from Speakeasy hosted MCPs
Tailscale operator · private ingressEnterprise-gated private MCP ingress through per-org Tailscale operator resources, workload-attested private serving, and per-server public/dual/private modes.
Externally hosted serversSupport proxying and collecting metadata on requests from MCP servers hosted outside of Gram
Observability, distribution, and governance around skill usage for an organization.
Persist per-tool MCP annotations so RBAC can make disposition-aware decisions in the remote-MCP proxy without fetching tools/list upstream on the hot path.
Employee compliance and token usage
CIMD lets OAuth clients use an HTTPS URL as their `client_id`. The authorization server fetches that URL on demand and treats the returned JSON document as the client's "registration." It removes pre-registration friction in the no-prior-relationship case
Make the existing global remote_session_issuer tier resolvable by tenants, curated by platform admins, and reachable by migration from duplicative org/project issuers.
End users see a logo and display name when challenged to authenticate against an external service, and administrators see that branding across the management UI. Requires expanding asset management to organization and platform tiers.
Observability and Reactions dashboard for admins
Given some risk finding in the dashboard that a user doesn't find useful they can select and suppress it such that it does not appear again in the future nor will it impact aggregations
A request-and-approve workflow for unverified MCP servers: evidence assembled automatically, an optional research agent, and a decision recorded with an explicit blast radius and a defensible rationale.
Non-enforcing replay of prompt-based policies over historical sessions + a review workbench, so authors can gauge efficacy and build a regression set before activation.
Derive policies from sampled activity (Policy Suggest), layer static rules ahead of the judge (hybrid), widen coverage beyond tool calls, and async session quarantine with tool blocking.
Using the RBAC system for policy audience assignment
Gives workspace admins visibility and control over Shadow MCP policy blocks — review, approve, or deny access requests and manage allow rules from the dashboard.
Allow users to specify the severity for findings identified by a given risk policy
New policy setting. In addition to flag and block we provide confirm for users to self approve flagged actions to unblock themselves. These are still logged to the admin for review.
Client detectionExpand Shadow MCP support to Shadow AI detection specifically client use detection, with device agent support as a prerequisite.
Every shadow serverExpand Shadow MCP request&approval to complete inventory of all shadow servers and analytics
Gives admins a choice to Allow All or Block All as the default disposition when creating a Shadow MCP policy, with block rules as the inverse of today's allow rules
Dashboard and Webhook powered request - approval flow for shadow server compliance
Handoff · same repoPick a captured agent session from the Speakeasy tracked agent sessions and continue it in another harness (Cursor, Codex, Claude Code).
Introduce a declarative endpoint for querying chat data, and all emitted metrics.
Usage · cost budgetsCustomers can bring their existing LiteLLM usage to the platform to track for Observability, Security and Cost budgeting.
Support multiple remote authorizations and standalone remote authorization client creation.
Create a scalable way to export customer data from AICP to customer-specified destination. Includes agent sessions and risk findings.
Customers will want cross-project management of certain remote session issuers, such as company-owned OAuth servers, to prevent duplication and simplify MCP Server management.