Skip to content

server@2.33.0

Choose a tag to compare

@gram-bot gram-bot released this 13 Nov 11:30
· 37 commits to main since this release
d5331fa

Minor Changes

  • c696162: Staff can see an organization's hooks installs on a new Hooks tab in admin and through Admin MCP: which versions each machine runs compared with the current release, how they were installed, which hooks endpoints they use, which machines have gone quiet, and how much traffic still reaches the legacy endpoints.

  • e46d4fd: Staff can manage platform GCP credentials from a new Platform credentials section in admin: list them, create one, view its configuration, run a live identity check, edit it and delete it. A platform credential's identity can no longer be changed, and it can no longer be deleted, while Speakeasy-managed organization encryption keys still reach KMS through it. Renames are still allowed. Provisioning a managed organization encryption key now re-checks the platform credential inside its write, so a credential deleted or repointed mid-provision is refused rather than recorded.

  • 8bc3776: Add a daily job that keeps MCP connection outcomes for 13 months: how each attempt ended and what fixed failures, per known provider, app and client, with no organization or person. A failure only an admin can fix counts an admin's change anywhere in the organization that can fix its reason, and is stored as unattributed when no such change or retry explains the fix. The first run backfills every whole day since connection events began.

  • 47b1fe9: Assistants no longer offer a "Set up agent identity" button, and the assistants.upgradeIdentity endpoint and the Platform MCP upgrade_assistant_workload_identity tool are removed. Speakeasy now moves each organization's assistants to their own agent identity: every assistant without one gets a dedicated agent, owned by the person who created the assistant, and its turns then run as that agent. The new agent can use every MCP server and skill in the assistant's project, and project inference when its owner can. An assistant whose creator has left the organization keeps running as before until it has a new owner. The organization's audit log shows these changes as made by the Speakeasy team.

  • c82c8e6: Speakeasy now moves organizations' assistants from the default runtime to the Kit runtime. An assistant is moved only once it has its own agent identity and is idle, and its next turns then run on Kit. The organization's audit log records each move as an assistant runtime change made by the Speakeasy team.

  • 43f4687: Connecting to an MCP server now requires an explicit mcp:connect grant, directly or through a role. mcp:read and mcp:write no longer imply it, so viewing or managing a server no longer lets a member call its tools. Blocks work as before: a block on a server still withdraws mcp:connect, and a grant made directly to a user that names the server still outranks a block inherited from a role or from everyone. Existing grants were backfilled with a matching mcp:connect beforehand, so no one loses access they had.

    Agent credentials issued before this change keep the mcp:connect their stored policy records next to mcp:read or mcp:write, and still need a live mcp:connect grant to use it.

  • bb9ed9b: Anthropic inference hooks now enforce security policies on tool call frames. With Validate tool calls on in Claude, each tool call Claude produces is scanned as a tool request with the same policies as prompt frames, and a block, warn, or quarantine match denies the frame before any of its tools run. Previously these frames were allowed unscanned, so a policy-violating tool call was only caught in the next prompt, after it had run. Tool call frames are not stored: the next prompt frame archives the same calls once. Tool-using responses now wait on the verdict within the same timeout as prompts, and a call that cannot be scanned in time is denied.

  • 61ec18f: The audit log page now opens on a Team changes view that hides Speakeasy's own background activity, such as publishing plugins and creating the API keys they use. That activity moved to a new Automated changes view, All changes shows everything, and the security view is now called Access. Older ?category=configuration links open the Team changes view. The auditlogs.list and auditlogs.listFacets endpoints take a new optional origin filter (user or system). You can also filter the list by date range and by resource: pick one from the Resource filter or use the filter button that appears beside a resource when you hover a row.

  • 02f2c41: Connection errors now show a reference id when Speakeasy expects to store a matching connection record, so people have an id to give their admin.

  • 721d768: Risk policies now act in real time in Claude Cowork. Hooks send prompt and tool events there, and the Stop transcript fills in what hooks miss, each message stored once. A warned prompt leads with a short approval link; warned tool calls get Claude's own permission prompt. Block messages mask the matched value.

  • 2c4c684: Custom domain setup now shows the domain's designated CNAME target and ownership TXT record. DNS configuration and infrastructure diagnostics appear separately, with one recheck action and independent DNS freshness. Staff can inspect target adoption and diagnostics in the admin dashboard and Staff Admin MCP; organization administrators can read saved diagnostics through Platform MCP.

  • bb329bd: Observe designated custom domain DNS targets deterministically. Verify ownership through TXT independently of routing, check infrastructure even when DNS is unconfirmed, stop automatically disabling domains after health failures, and hold missing or not-ready certificate alerts until the domain's DNS points at Speakeasy.

  • 5e45770: Add an "MCP connection logs" data export. When enabled, failed MCP connections and server problems an admin or service provider needs to fix are sent to the project's OTEL destination as speakeasy.mcp_connection events. By default, blocked or declined connections are not included. Connections through organization gateways are not included.

  • 0b95ea6: Organization admins can import Gemini web chats into Agent Sessions through Google Vault. Connecting takes two saves: paste the service account key and enter a delegated Workspace admin, then pick the organizational unit to export, or Entire organization. Speakeasy finds or creates its own Vault matter, and syncs completed UTC days every hour. When Google refuses access, the error names the delegation scope or Vault privilege to fix.

  • 8b1ec11: Setting up a personal sign-in for an MCP server now always creates a sign-in of the server's own. The dashboard offers New from scratch or New from existing in place of reusing a stored one: starting from an existing sign-in copies its settings for review into an independent sign-in, and the stored client secret is copied on the server without being shown. This stops a member's sign-in for one server from replacing the access another server obtained through a shared sign-in. Service accounts and stored or caller-supplied credentials can still be shared between servers, and existing shared sign-ins are left as they are.

  • a87c444: MCP connection log exports now have a minimum level. The default stays at error, which sends failed connections. Set it to warn to also send blocked connections, or to info to also send declined ones. Set it on the export, through the API, or with the Platform MCP create_data_export tool.

  • 2fa5e69: Masked chat summaries for Organizational Memory are now written by the chat analysis pipeline instead of a separate 15-minute sweep. They run for organizations with Organizational Memory turned on (the Business Memory Extraction toggle in the staff admin dashboard), after a chat has been quiet for 30 minutes, and again each time a chat resumes and goes quiet. Fact extraction now reads masked transcripts and masks the facts it stores.

  • 765f9d6: Adds the orgEncryption management API for organizations with customer-managed encryption keys. Admins can register a GCP KMS symmetric key, verify it, switch the organization to it, complete or roll back the switch, rotate a use case's data key and resume after a suspension. A key that still protects data cannot be deleted, and Speakeasy-managed keys cannot be deleted at all. The externalKeys API now only ever returns signing keys, and a JSON Web Key Set refuses an encryption key.

  • edf7d08: Members without log:read can now read the tool-call logs and traces of their own sessions and of the agents they own. The four raw log endpoints narrow to that activity instead of refusing, so an agent's owner can troubleshoot it without a grant that reveals every member's tool arguments and results.

  • a87c444: Platform MCP adds set_data_export_min_severity, which changes the level of an existing MCP connection logs export after confirmation. Audit records for creating, pausing and resuming these exports now include the level.

  • a7a3dea: Platform MCP members now see their own activity without organization-wide permissions. The list_chats, search_chats and read_chat tools cover the caller's own chats and the chats of the agents they currently own, and the search_tool_calls and list_recent_tool_calls tools cover the tool calls of their own sessions and their agents. A chat:read or log:read grant on the project still widens each tool to the whole project, and no filter widens a member's view.

  • d733ce1: The request access page now shows which server and tool you are asking for: the server name, its URL or, without one, the identifier the agent used, and the tool. A new risk.getPolicyBypassRequestTarget endpoint reads a request link without filing the request.

  • 1a47bde: Members can now reveal the matched value of a risk finding on their own sessions, and on sessions of agents they own, without holding the organization-wide chat:read permission. The reveal follows the same ownership rule the session view already uses and is audited the same way.

  • 57aea4e: Session lists now follow chat:read. telemetry.listSessions lists per-session rows only for projects the caller holds chat:read on and otherwise reports the slice's totals alone, flagged with scoped: true. telemetry.searchChats narrows to the caller's own sessions and those of agents they own instead of listing every member's.

Patch Changes

  • af2eebf: Staff can list Organizational Memory eval runs and read their scores through Admin MCP.

  • 593ae16: Staff can start Organizational Memory eval runs against the seeded demo corpus and list or read their results through a new admin API. Runs go one at a time, are capped per rolling 24 hours, and a run left unresolved past any workflow's lifetime no longer blocks later starts.

  • fe585ec: The staff admin "Business Memory Extraction" control is now an "Organizational Memory" on/off switch with no cap input. Switching it on keeps a stored daily cap or sets the default of 100, and an organization whose cap is 0 now counts as off for Organizational Memory instead of showing the team memory tools while nothing gets summarized.

  • 7fcc94e: The mcp.connection.steps metric now counts only connection rows stored in ClickHouse, so organizations with logs off no longer appear in it. Each row that reached a provider is tagged with that provider when it is a known vendor, otherwise other.

  • 7fcc94e: Staff can see whether a known sign-in provider is failing across organizations with the Admin MCP describe_provider_health tool. It reports counts only and names no organization.

  • 3dcf138: Assert worker schedule registration against the schedules the Temporal client created rather than a hand-maintained count polled from the eventually consistent schedule listing, so a new schedule can no longer leave the test passing on some runs and failing on others.

  • 1b87527: Allow Channels OAuth consent submissions to complete browser redirects to registered client callbacks using the callback-aware form-action allowlist shared with Platform MCP.

  • 3147ef7: Fix Channels OAuth consent submissions being rejected by the browser Origin check, and style the consent screen with the shared Speakeasy authorization layout.

  • 014f4a3: Platform MCP chat search now masks emails, phone numbers, card numbers, SSNs, medical record numbers, URL passwords and one-time codes in transcripts, not only the credentials the secret scanner finds.

  • 883bb93: The View agent sessions and Manage agent sessions permission descriptions now say that every member keeps their own sessions and those of agents they own, matching what the Agent Sessions list already shows.

  • 3ac425d: Cloning an MCP server now requires MCP access in the server's own project. A role whose MCP access is limited to a different project can no longer clone servers outside it.

  • 8a32344: Connection history fixes:

    • Token requests from Claude Code, Claude, ChatGPT and Codex now show the app's name instead of "other". These apps send a generic User-Agent, so Speakeasy names them from their verified client ID. A script that borrows a local app's client ID still shows as a script, and a client that names itself something else shows as "other".
    • A tunneled gateway member that times out after its server started answering is now recorded as unreachable instead of as a Speakeasy internal error. A timeout before the tunnel passed on any answer is still recorded as a Speakeasy internal error.
    • Rows written later for held-back repeat requests keep their request's trace id.
    • A failed provider refresh no longer waits up to 5 seconds on a binding lookup.
    • On shutdown, Speakeasy finishes writing connection rows before ClickHouse closes, and counts rows it could not write as dropped.
  • 624e5ff: Demo and local seed data now include a member who owns an agent, with agent sessions and tool logs, to show owner-scoped inspection.

  • f5955e3: Drop the unused oauth_proxy_servers, oauth_proxy_providers and oauth_proxy_client_info tables.

  • 8fc751d: For organizations enforcing typed auth clients, an assistant's credential for each MCP server is now chosen from the auth clients the server offers, so a server offering one usable client no longer waits for an administrator to pick it. An assistant turn whose credential choices cannot be loaded now fails and is retried instead of running with every connected server unavailable.

  • f087bc9: The upstream OAuth migration no longer reports a server as needing a client configured only because another server on the same upstream URL has its own OAuth client with active grants.

  • 973381b: Allow gateways to use and select live sign-in issuers shared by their organization, so settings remain editable after issuer consolidation. Keep automatic refresh working for project-owned OAuth registrations used through organization-shared sign-ins. Issuers from other organizations or other projects remain unavailable.

  • fbc465d: The first Gemini sync now imports the last 30 days in one Google Vault export that runs up to the time the sync starts, instead of one UTC day per export starting with the oldest. While an export runs, Speakeasy checks it every 5 minutes, so conversations appear soon after Vault finishes rather than over the next two and a half days. After that, each hourly sync exports everything since the previous export, so today's conversations no longer wait for the UTC day to end. An export too large to import is split into shorter time ranges. Changing the service account key or the organizational unit restarts the 30 days from that save, not from when Gemini was first connected. Connections already syncing day by day continue from where they are, up to the present, on their next sync.

  • ec5739c: Serve the hooks@0.3.52 binary to hook installations. Previously pinned releases stay available so installations that have not regenerated their bootstrap script can still install.

  • 6a8807d: When an AI tool runs hooks with HOME pointed at a temporary folder, the hooks binary now finds the device agent under the account's real home and asks the running agent before any machine-wide copy. It no longer runs an outdated machine-wide agent that could open a stack of macOS "Keychain Not Found" dialogs.

  • ce95ba1: Keep typed auth clients and credential choices synchronized when consolidating authorization servers, without requiring a separate upstream OAuth migration run. Existing explicit choices and permissions are preserved. When both authorization servers had a client for the same MCP server, existing sign-ins keep the client they were already using.

  • eb5ad7b: Add a table for settings that only MCP connection log exports have, starting with the lowest log level an export sends. Nothing reads it yet.

  • d9f2b98: The upstream OAuth migration page no longer reports an organization as not migrated because of connections made through a gateway, which only ever need choices for the servers approved when connecting. Once enforcement is on, assistant profiles that no turn has used yet are shown separately instead of as work left to record.

  • 3617ddf: Adds per-organization envelope encryption for new features: each organization gets its own key encryption key in KMS, minted by Speakeasy on first use or supplied by the customer, wrapping one data encryption key per use case. An hourly sweep probes those keys, suspends organizations whose key became unusable, rolls back a key switch whose new key stopped working, and disables Speakeasy-managed keys a week after nothing uses them. It stays off until GRAM_ORG_ENCRYPTION_CREDENTIAL_ID is set, and no existing data moves to it.

  • e9c9384: The Project Assistant can use its platform tools again after it is given an agent identity. When a person sends it a message, these tools now act for that person, with that person's own access, exactly as if they had called the tool themselves. A turn that acts for nobody, such as a scheduled trigger, gets a clear answer that the tools need a person, instead of a failed call it keeps retrying.

  • b631ed0: Connection health now records three more kinds of failed connection. Each record says what failed and who needs to fix it. Apart from the inference token change below, what the client receives is unchanged.

    • When a workload or an assistant asks Speakeasy for a token, the request is recorded, whether it worked or not. This includes the shared sign-in address that serves all of an issuer's servers. Refused requests from one workload issuer are grouped, so tokens with new names can't flood the history. A grouped record that stands for several requests doesn't name an identity. A workload's request for an inference token now checks its assertion before its project, as the platform token request does, so a request for a project outside the issuer is recorded too. An invalid assertion is now refused as invalid even when the project is also wrong.
    • When a hosted server's tool call is refused by the service with a 401, or a 403 that says the token is invalid, the record says whose credential was refused: the person's sign-in, the organization's shared sign-in, or a saved key. Nothing is recorded when the caller supplied the credential or chose the server's address.
    • When a direct server rejects the sign-in on initialize, tools/list, or any other request, not only on tool calls, the record names the request. A rejection counts only when the request that reached the server carried the credential, and a 403 counts only when it says the token is invalid. Servers behind a gateway follow the same rules. A failed renewal of a server's own credential during a provider outage is recorded as that outage.
  • 7550c88: Remove intent decisions from the command palette and retire the launcher judging endpoint. Command palette results now use local fuzzy ranking only.

  • 101d594: Staff now manage platform credentials only through the admin dashboard. The older staff-only /rpc/adminExternalCredentials.* routes, which accepted dashboard sessions, are removed.

  • 21e76ce: A member's access listing no longer shows a skill they are blocked from reading just because they also have permission to edit skills. A read block now hides the skill there, as it already did everywhere else.

  • 15687e1: The Platform MCP has a new remember tool for organizations with Organizational Memory turned on in the staff admin dashboard. It saves a decision, convention or gotcha to a project's Organizational Memory. Detected secrets are masked before the memo is stored, the memo is attributed to whoever saved it, and every save is audited. It needs memory:write on the project: organization admins have it by default, members need a grant.

  • 18de7bb: Add the memory:read and memory:write permissions for upcoming team memory in the Platform MCP. memory:read lets someone search and recall a project's team memory, and memory:write lets them save memos to it. The built-in Admin role holds both by default. Grant them to members per project through a custom role, on the Access page or, for organization admins, with the new Platform MCP configure_memory_access_role tool.

  • 15687e1: The Platform MCP has a new search_team_memory tool for organizations with Organizational Memory turned on in the staff admin dashboard. It searches one project's Organizational Memory, best match first, and covers masked session summaries, saved memos, and day, week and month digests. Every hit carries the source to cite. Sessions that are excluded from memory, deleted or from personal accounts never appear, and neither do memos saved from them. It needs memory:read on the project (organization admins have it by default, members need a grant), and every search is audited.

  • 8895622: The MCP server & tool insights page and the Tool logs page load faster and no longer time out on long date ranges. Splitting skill loads by how they arrived (hook or MCP) now reads pre-aggregated trace data instead of scanning every raw log in the range.

  • 3a1d8fd: In an organization that enforces upstream auth, Speakeasy now decides a connection's upstream choices from the auth clients each server offers, and no longer from the older upstream OAuth client associations. A server whose OAuth offer was removed now works without authentication even when an old association is left behind, and a client offered to a server is recorded for it even when no association was ever made. This applies to new sign-ins, dashboard sessions, gateway reviews, and gateway discovery. Automatic choices still only use clients bound to the session's sign-in issuer. The upstream OAuth migration report and the choice backfill read the same state, so on an enforcing organization they no longer report these connections as gaps.

  • a9a3fa6: When Speakeasy can't read a person's saved upstream access token, it now renews the sign-in with the saved refresh token during the call, without the person noticing. When nothing saved can be read, the connection shows as not connected, so connecting again replaces it: direct servers send the usual sign-in challenge, and gateways ask the person to reconnect that server instead of calling it without credentials. These requests are recorded as stored_sign_in_unreadable.

  • c8d902f: Admitting a workload now refuses a malformed wimse:// or spiffe:// subject: it needs a trust domain and no user information, port, query or fragment, so a rule that could never match is caught when it is saved. Withdrawing an admitted subject now returns withdrawn_tier, which says whether the organization or project admission was removed.