You can also browse, search and filter all release notes directly in the application → Open release notes in TAI.
Connectors: build your own, no code required
Major New Feature. Only Company Admins.
The Connector Builder and the Connector Lab are now open to every company admin. Until now, connecting an agent to a system TAI did not ship a connector for meant asking Nezasa. From today you open Connectors, click Create connector, and describe the source in a chat. The Builder does the rest, and a verified connector is ready for your agents in one sitting.
What you can connect
REST APIs: public data APIs, the SaaS tools your company already uses, internal services.
GraphQL APIs: one endpoint, read-only queries.
SQL databases: PostgreSQL and MySQL, one SELECT at a time.
NoSQL databases: MongoDB, with find and aggregate reads.
MCP servers: every tool the server offers becomes an operation.
Feeds and web pages: a JSON or XML feed, or a single HTML page.
How a build goes
Describe, then approve a plan. Say what your agents should be able to answer. The assistant proposes the hosts, the credential fields and a curated set of read-only operations, and generates nothing until you approve.
Self-tested against the real source. The assistant writes the connector, runs every operation against your live system and fixes failures itself.
Verified before it goes live. An independent check of the contract, the hosts reached, the code and the stated intent. You read the report and click Activate.
Credentials never enter the chat. Enter values in the Credentials panel; TAI attaches them at call time, outside the connector's sandbox. Paste a key into the chat and the assistant refuses it.
Read-only by construction. No connector can create, update or delete anything in your systems, whatever you ask for.
Test drive in the Connector Lab. Ask a live question, watch every call the connector makes, then assign it to an agent.
Yours to keep, rename or retire. A connector your company built is edited and deleted from its detail page, like any agent.
Browsing the connector catalogue stays open to all users. Building, editing, deleting and test-driving need the Company Admin role. Learn more
One chat, six kinds of source: your agents call a connector built in the Builder, which reaches REST and GraphQL APIs, SQL and NoSQL databases, MCP servers and feeds through the TAI gateway. Read-only, credentials outside the sandbox, only the declared hosts.
Create connector is now on the Connectors page for every company admin. A name and a one-line description are all it takes to land in the Builder.
One sentence in, a plan out. Describe the source and the questions agents should answer, and the assistant checks the API itself, then proposes the host it will reach and a table of operations for you to approve.
For a source that needs a key, the assistant declares the credential fields and asks you to fill them in the Credentials panel on the right. The value is stored encrypted and never appears in the chat.
The assistant narrates its self-tests as it goes. Here the live API first rejected the key, the assistant asked for it to be checked, then both operations passed and the connector went to verification.
After approval the assistant writes the connector and self-tests every operation against the live source. The Submitted connector panel shows the verification report: egress, intent, code safety and contract, with findings. One click on Activate and the connector is live.
Test drive in the Connector Lab: ask a live question, get the answer as a table, and watch every call the connector made in the trace on the right, with raw and shaped byte counts.
Chats now screen every turn for injection, secret leaks and cross-tenant reach
New Feature.
Chats in the Workbench and in an embedded TAI chat are now screened by built-in security guardrails.
Messages that try to override an agent's instructions, get it to reveal how it is configured, or reach into another company's data are refused — you get a short notice instead of an answer.
Secrets and payment-card numbers are removed from replies as they are written.
What a connector or a web page brings back is treated as data, so instructions hidden inside it are not followed.
A refused exchange is left out of the rest of the conversation, and every block and redaction is recorded. The Agent REST API is not screened yet. Learn more.
![A chat reply showing an AWS credentials snippet in which the access key id reads [REDACTED].](https://nezasa.intercom-attachments.eu/i/o/w8jlyl9x/182890760/aed6b529218b197bbe19b66a34ea/01-output-redaction.png?expires=1792384200&signature=8d023dcc963a6b140db17f7830b9bb51c91caa5176fdd7986580739cb5e27d5f&req=0d1tzVH6qzEolxv889osp%2BKi7L61WkTSfvmVGQ%2Fat8HGVAL0uRvt%2F8NXpEWr%0AYBA3dI02V%2FLdbWQeEA%3D%3D%0A)
A secret in a reply is replaced with [REDACTED] as the answer is written; the surrounding text is untouched.
A message that tries to override the agent's instructions gets a short notice instead of an answer.
Give your agents a bigger context budget for long runs
New Feature. Only Company Admins.
Settings → Context budget shows the token allocation Nezasa has granted your company for conversation history (the platform's standard budget unless Nezasa has granted you more), and lets you set your own default below it — a higher budget means an agent remembers more of a long conversation or a large report run before older turns are trimmed, at the cost of re-sending more history (and spending more credits) on every turn. Every agent is listed on the same screen, so you can give one its own budget instead of the company default. Trip Guard, being built in, runs on a 300,000-token budget set by Nezasa — not editable per company and not limited by your allocation — so a full multi-pillar risk review keeps its earlier findings in view for the whole run. See Context budget.
Company admins now see advisor runs as their own Analytics channel
Improvement. Only Company Admins.
Advisor-run activity used to land in the REST API bucket of every by-channel breakdown, inflating it with spend that never came from a real REST API call. It now has its own Advisor Run channel, so REST API figures reflect genuine API traffic and advisor spend is still visible, just correctly labelled. Past periods have been relabelled too, so a previously exported REST API figure may now read lower.
Agents: Claude Sonnet 5 is now the default model
Improvement.
TAI's built-in agents, and any agent that leaves the model unset, now run on Claude Sonnet 5 — Anthropic's latest Sonnet, with sharper reasoning and more reliable tool use at the same responsive speed. Nothing to configure: existing agents pick it up on their next conversation. Agents pinned to a specific model keep it.
Prompt Advisor flags prompts that repeat content in two formats
Improvement.
The Prompt Advisor's Review flags a prompt that asks for the same content twice — a chat recap of a file it already wrote, or a table restated as an unconditional export — under a new Output economy finding category.
Passolution stops repeating itself on quick lookups
Improvement.
A single-destination entry-requirement lookup used to answer, then restate the same facts under a second Summary heading. Passolution now answers once — quicker to read, and quicker to act on.
Itinerary checks now cover the rental car: fit, child seats, driver
Improvement.
Vehicle fit: a compact booked for a safari or gravel route, too few seats for the party, a manual for a driver from an automatic-only market, missing winter kit where it is an optional extra.
Children: a child, booster, or infant seat the child's age requires but not booked as an extra — or one seat for two children.
Driver: young- or senior-driver limits from the driver's date of birth, an IDP reminder, a named driver who isn't a passenger.
Timing and borders: pickup too close to landing, drop-off too close to departure, uncovered rental days, cross-border or ferry rentals without permission.
Each check runs only when the car booking carries the detail it needs — a thin booking raises nothing. The one exception is the driving-permit reminder, which needs only the driver's nationality and the destination.
Itinerary checks now flag a tour pickup or ferry you can't reach
Improvement.
The Itinerary Checker now treats a tour pickup, a ferry sailing and its check-in cutoff, a scheduled transfer or a timed entry as a hard deadline — unlike a hotel night or a rental-car day, which have no single moment to hit.
A late drive, checkout or connection upstream is reported against the service you would actually miss, not against the one running late — so it is clear what to rebook.
Where the booking doesn't give a start time or a cutoff, the check says so rather than guessing a deadline.
See Itinerary Checker.
Opening a deep link while signed out now lands you back on that page
Fix.
Opening a deep link — a specific agent, report, connector, or any other page — while signed out correctly showed the login screen, but signing in always dropped you on the dashboard instead of the page you were trying to reach. Signing in (password, MFA, the local bypass, or Google) now returns you to the original link, including a session that expires mid-use while you're on a page.
Chats: a withheld answer now disappears when a guardrail blocks it
Fix.
The blocked answer used to stay on screen with the notice added below it, until you reloaded. The notice now replaces it, along with the reasoning and any files the turn produced.
SDK integrators: ships in @nezasa/tai-chat.
A few already-read release notes may briefly show as unread
Fix.
One-time cleanup after this release. Opening Release Notes clears them — nothing else to do.
Settings → Credentials loads even if one connector's saved contract is broken
Fix. Only Company Admins.
A broken saved contract on one connector used to fail the whole list; it now shows as broken, and its credential can still be removed.
Reports: HTML reports are no longer repeated in the chat reply
Fix.
When an agent produces a downloadable HTML report, the chat now ends with a short summary and the download instead of repeating the whole document. Report runs finish faster and use fewer credits.
Reports: a run's failure reason now reads in your language
Fix.
The reason on a Failed or Incomplete run was stuck in whatever language the run was started in — so switching your Workbench language never changed it, and a colleague's run could come back in theirs. It now follows your own language: on the run detail card, in the library's hover text, in the scheduled-report failure email, and on most notices a system-paused schedule shows. Older runs are translated too, though a few of the oldest — those recorded in another language, or giving a reason since reworded — keep their original wording.