Changelog

Per-project guardrails

Security

Prompt Injection Protection and DLP used to apply the same way across every project in an organization, whether or not that project needed them. Guardrails are now configurable per project, so you can turn on Prompt Injection Protection and set DLP rules for the projects handling sensitive data without changing behavior everywhere else. A project running internal tooling can stay permissive while one handling customer data gets stricter rules.

Web search tool support for Parallel

New

Models routed through Gateway couldn't reach the open web without you wiring up a separate search tool. Gateway now supports Parallel Web Systems as a web search tool, so models can issue search queries and get results back as part of the same tool-calling flow you already use for everything else.

Multi-modal support extends to audio and video

Improved

Multi-modal support in Gateway centered on text and images, leaving audio and video models without full support for their inputs and outputs. That support now extends to audio and video-based models, so you can route requests to models built for those formats through the same Gateway interface.

GitHub connector expands to 145 tools

New

Agents working against GitHub previously had to fall back to raw API calls for anything outside basic issues and pull requests. The GitHub connector now covers 145 tools, adding sub-issues, milestones, assignees, releases, workflow re-runs and artifacts, pull request review threads and diffs, commit comparison, branch protection, and discussions. Request latency is now bounded and the retry behavior that could duplicate issues, comments, and releases has been removed, so higher-volume workflows run more predictably.

Outlook payload sizes cut

Improved

Every Outlook message and calendar tool call returned full HTML bodies and quoted reply chains by default, even when an agent only needed the latest reply. Outlook tools now return plaintext bodies by default, trim bodies out of list results, and support an option to drop quoted reply history. A single message dropped from roughly 57kb to 14kb, and a ten message list from 394kb to 16kb; calendar tools kept an HTML opt-in for agents that still need rich formatting.

Axiom MCP Server

New

Teams running Axiom for log and event data had one path in: the first-party Axiom connector. Axiom MCP Server is now available as a hosted remote option alongside it, so agents can reach Axiom directly over MCP without you standing up and maintaining your own server.

Usage history for enterprise plans

New

Enterprise admins had no way to see how credit consumption changed over time, only a point-in-time balance. Billing now includes usage history, giving admins a running record of consumption over time. That makes it easier to catch a usage spike early or plan ahead of a renewal.

SharePoint sensitivity labels

File Storage [[hide]]

New

Files and folders synced from SharePoint carry Microsoft Purview sensitivity labels, but that classification data never made it into Merge. GET /files and GET /folders now return each item's assigned sensitivity label, so you can see how content is classified without a separate lookup. Content assigned a label you've excluded is left out of syncs entirely, and gets removed automatically if a label changes after the fact.

Accounting sync controls tighten

Accounting [[hide]]

New

Contact syncs from NetSuite pulled customers and vendors together, forcing you to filter them apart downstream. Selective Sync now supports filtering GET /contacts by contact type, so you can exclude either customers or vendors directly at the source. Xero's connection flow also picked up support for granular OAuth scopes, so new Xero accounts that were failing to link can now authorize successfully.

HRIS fills in employment and time off data

HRIS [[hide]]

Improved

Employment records pulled from SAP SuccessFactors only ever showed an employee's current role, even when a promotion or compensation change was already scheduled. GET /employments now returns both current and future-dated employment records, so upcoming changes show up before they take effect. Sage HR's GET /time-off endpoint also now populates full-day requests correctly instead of returning empty amounts, and Workday syncs no longer mark active employees as deleted.

Kimi-K3, Makora, and Modal added

New

Merge Gateway added three new vendors this week: Kimi-K3 through US-based inference providers, Makora, and Modal. All three ship with zero data retention, so requests routed through them aren't stored by the provider after the response returns. Check the model catalog for context windows and pricing before pointing existing Gateway requests at any of the three.

Deepseek V4 Flash upgraded

Improved

Deepseek V4 Flash on Merge Gateway is upgraded to the latest official release from Deepseek. Existing requests to the model pick up the new version automatically, so no action is needed on your end.

Hundreds of MCP connectors join the catalog

New

Wiring up a new tool for your agents used to mean standing up a separate MCP server or building an integration from scratch, even for connectors Merge hadn't built itself. Hundreds of generic MCP connectors are now browsable and enabled directly in the Agent Handler catalog, running on the same authentication, access control, and governance infrastructure as every Merge-built connector. There's no server to host and no integration to write: enable one from the catalog and it inherits the same guardrails as the rest of your stack.

Seven new connectors, plus webhooks for Teams and Zoom

New

Seven new connectors landed in the catalog this week: Okta, Brex, Microsoft Excel, Metabase, Sigma Computing, Statsig, and Granola. Microsoft Teams and Zoom also gained webhook support, so agents can react to activity in a channel or meeting as it happens instead of polling for it. Between the new connectors and the webhook support, agents now cover identity, spend, spreadsheets, BI, feature flags, and meeting notes without a custom build for any of them.

AI Guardrails

Security

Running agents against real customer data meant either trusting the model not to leak anything sensitive or building a scanning layer in front of every tool call yourself. The new AI Guardrails experience adds PII detection, a guardrail rule table for configuring what gets allowed, redacted, or blocked, and a live tester that shows each check running in real time. You can see exactly which rule fired on a given call instead of finding out after the fact.

Gmail and Google Drive hardened

Security

Gmail drafts could previously lose data or send twice when a request was retried, and the connector's OAuth scope was broader than what it actually used. Both are fixed: drafts save reliably, retries no longer duplicate a send, and the OAuth scope is narrowed to only what the connector needs. Google Drive got the same treatment: shared-drive writes are dependable, export errors are clearer, file reads return plain text, and PII no longer ends up in logs.

FactSet connector retired

Deprecated

The FactSet connector is retired and no longer available in the Agent Handler catalog. If your agents relied on it, check the connector catalog for an alternative.

Dashboard Scopes page redesigned

New

Configuring what data Merge can read or write from a linked account used to mean opening a modal for every object and field, one save at a time. The Scopes page in the dashboard is redesigned: search, sort, and filter across every object and field, and edit read and write access for multiple objects in a single save. Read and write toggles are now visually distinct at a glance, so reviewing a scope configuration no longer means opening each row to check it.

Hive support added to Ticketing

Ticketing [[hide]]

New

Teams running support on Hive had no way to bring ticket data into Merge. GET /tickets, GET /comments, and GET /attachments are now supported for Hive, syncing through the same Ticketing API used across every other help desk integration. No separate mapping or custom logic needed to bring Hive tickets, comments, and attachments into your app.

Freshservice syncs go real-time

Ticketing [[hide]]

New

Freshservice data previously updated on a polling schedule, so a new comment or tag could take a full sync cycle to show up. Real-time webhook syncing is now supported for GET /tickets, GET /comments, GET /attachments, and GET /tags on Freshservice, so ticket activity reaches your app as it happens instead of on the next poll.

Merge Fusion

New

Every prompt used to cost frontier prices, whether or not the task actually needed frontier-level reasoning. Merge Fusion sends one prompt to a panel of models in parallel, then has a judge model synthesize their answers into a single response, with you choosing both the panel and the judge. On the DRACO benchmark, an all open-source panel matched Fable 5 at a quarter of the cost, and a premium panel beat Fable 5 outright by 8.5 points, the new best score on the benchmark. Set "model": "fusion" and pass in your panel to start using it.

Indirect prompt injection protection

Security

Content pulled into a prompt from a web page, document, or tool output can carry instructions that quietly redirect a model, and Gateway had no dedicated way to catch that pattern. Indirect prompt injection protection is now live, with per-axis mode and sensitivity controls so you can tune how aggressively it flags content on each axis. A built-in detection tester in Security settings lets you check how a specific payload would be scored before you rely on it in production.

Flex and priority service tiers

New

Every request ran at the same priority and price, whether it was latency-sensitive or could wait behind other traffic. Flex and priority service tiers add a service_tier request parameter, with billing and per-tier pricing on /v1/models reflecting whichever tier you choose. Flex costs less for work that can tolerate queuing; priority costs more and jumps ahead of it.

Google Ads connector

New

Agents working with paid search had no direct way to read or adjust Google Ads campaigns without building against Google's API and GAQL syntax themselves. The Google Ads connector adds 22 tools spanning customers, campaigns, ad groups, ads, budgets, and keywords, plus a dedicated GAQL search tool for custom queries. Agents can now pull performance data or adjust budgets and keywords through the same normalized interface as every other connector.

HubSpot ticketing

New

HubSpot connections only exposed CRM tools, so agents handling support tickets in HubSpot needed a separate connector or direct API access. Ticketing is now available as an optional scope: create, read, update, delete, and search tools for HubSpot tickets. Turning it on doesn't change anything about an existing CRM-only connection, so teams that only want contacts and deals keep the same scope they had.

Generic MCP Configure wizard

New

Bringing a Generic MCP catalog server online used to mean stepping outside the product to register a client and collect credentials before it could be added. The new Configure wizard handles import and credentialing for a catalog server on the same page, including automatic client registration when the server doesn't require a client ID upfront. Setup that used to span multiple tools now happens in one flow.

Per-integration Link setup

New

Every integration in Link showed end users the same auth and Selective Sync screens, even when a step didn't apply to that integration. You can now hide individual authentication and configuration steps for a specific integration, and turn off the Selective Sync screen for specific integrations while it keeps running in the background. End users see fewer screens and connect faster, without losing the syncing behavior on your side.

Sync status explains itself

Improved

A model stuck in SYNCING gave no indication of whether it was actually behind or just waiting its turn. GET /sync-status now returns sync_status_reason, set to RATE_LIMITED or WAITING_ON_OTHER_MODELS when a model isn't progressing, and data_fresh_as_of, the timestamp through which that model's data is current. Both fields are live in Python SDK 4.1.0 across every category's SyncStatus object.

Sync reliability hardened across HRIS and File Storage

HRIS and File Storage [[hide]]

Improved

A SharePoint sync interrupted partway through used to start over from the beginning, and Workday's leave-of-absence records had nowhere to map on the TimeOff model. Large SharePoint and OneDrive syncs now resume automatically where they left off, and Workday leave-of-absence records map to a new LEAVE time-off type. SAP SuccessFactors rehires get their termination dates cleared and active status restored automatically, and unchanged employee records stop getting flagged as modified on every sync.

Embedded Routing Stack

New

Giving every customer their own model controls used to mean building tenant isolation and per-customer routing policy on top of Gateway yourself. Embedded Routing Stack adds that layer to your existing setup: provision tenants, set routing policies per customer, and support bring-your-own LLM keys, all from your API. It works inside your current Gateway setup, so there's no migration and nothing else changes.

Sail Research and Xiaomi Mimo added

New

Two new model vendors joined the catalog this week. Sail Research and Xiaomi Mimo are now available as model vendors on Merge Gateway, giving you more options to route to without building a new provider integration yourself.

Better support for coding agents

Improved

Routing model calls from a coding agent through Gateway previously meant extra friction configuring the base URL and headers for each tool. Merge Gateway now has enhanced support for coding agents such as Claude Code, Codex, and OpenCode, so pointing those tools at Gateway and switching the underlying model is simpler.