October 2026 product release 

No tricks this October season, but lots of product treats for you! 

Barndoor Automations now has its own design and build agents: a design agent reviews your request as a plan, and a build agent runs it once you approve. We’re also rolling out full visibility into MCP tool changes and what they mean for your policies, agent-scoped data controls down to the field level, policy-based LLM routing, and native governance for decision models starting with Jev, a classifier built for fast, cheap routing decisions.

Let’s dig in!

Review-before-build in Barndoor Automations 

Most AI automations tools give you a starting point, but leave a gap. These tools will create a first pass at the workflow or route a task to a human, forcing you to fix or patch the logic and check if it completed accurately. Barndoor Automations carry your request to a working automation, so you can approach workflow automation with confidence and control. 

To achieve this, Barndoor is launching two dedicated agents for building AI automations (we call them Blueprints): a design agent that turns your request into a Blueprint, and a build agent that executes it once you approve the plan. 

The design agent turns what you asked for, in plain English, into a “brief” which contains your requirements,  then a “sketch” where every step traces back to a requirement. Once you sign off, the build agent takes that approved sketch, writes the automation, and checks its own output against what you approved. Every Blueprint inherits the access, data protection, and spend controls you’ve already set up; no separate policy to configure.

Build your first automation with Barndoor’s design and build agents, starting today.

AI Gateway updates

Deeper MCP change tracking. MCP servers often change without warning, which impacts your enterprise access policies. Barndoor now checks every MCP server on a schedule, which can be triggered on demand as well. When a tool listing changes, Barndoor saves a new version and shows the differences across descriptions, operations, annotations, and input and output schemas. Each change is traced to the access policies related to the affected tools. In the admin control center, admins can filter to policies with tool changes to review the before and after for each tool before deciding whether a rule still holds. 

Context-aware DLP for AI traffic. A new policy category for conditional policies, available first for Google Drive and Slack, which are enforced based on a resource’s context vs. merely its contents. 

  • For Google Drive, conditions key off a file’s metadata: its label or name, matched as a literal, wildcard, or full regex. For example, you can block anything labeled “Confidential,” redact PII on files matching payroll, or catch an entire naming convention like ^FY2[0-9]board meeting.* with a single pattern. 
  • For Slack, conditions key off where a conversation happens: channel name (literal or wildcard) or channel type, so private channels and CMs can carry stricter rules than public ones.
  • Uncached meta data, meaning it’s checked fresh every time an agent makes a request. If someone renames a file to match a blocked pattern, or relabels it as “Confidential,” or a Slack channel gets switched to private, the new policy applies immediately on the next request so there’s no lag where the system is still working off stale information.
  • Net effect for an admin: policies can follow the organizational structure that already exists (labels, naming conventions, channel types) instead of requiring content scanning for every case, and every decision is captured in the audit log.

System 1 models are now natively governed, starting with Jev

A growing share of AI models aren’t chat models at all, rather they’re classifiers that return a probability or a bucket instead of generated text, useful for things like model routing. 

Barndoor AI Gateway now governs decision models, starting with Jev, TypeSafe’s System 1 model. Jev plugs in two ways: call it directly as a model, with the same policy enforcement and logging as any chat model, or use it as the decision-maker inside a routing policy, where it answers the yes/no classification the router needs. In our own tests, Jev matched Claude Haiku’s routing decision quality while running roughly 7x faster at a fraction of the cost, and we log and policy-enforce every call like the rest of your AI traffic. Read our CTO’s take on this. 

New MCP connectors this month 

Typeform: The Typeform MCP server lets agents build a form, publish it, and then read the results. 

Mixpanel: The Mixpanel MCP server lets agents answer product questions against live event data and save the result as a report or dashboard, instead of hand-building charts.

SocialSight: The SocialSight MCP server lets agents turn a brief into finished marketing creative such as ads, UGC-style videos, faceless content, and social post visuals,  by choosing the right model and submitting the generation job in one step.

Gusto: The Gusto MCP server lets agents answer payroll and people questions straight from the system of record. 

Socket: The Socket MCP server lets agents check a dependency’s security and quality score before adding it, so risky or typosquatted packages get caught while code is being written rather than in review.

Sign up for a demo

Ready to empower your knowledge workers with governed AI at scale? Book a demo or take a product tour to see Barndoor in action.

Note to customers: If any of the features above aren’t visible in your account yet, reach out to your account executive and we’ll get them turned on for you.