kaiva BRIDGE
CHANGELOG

What shipped

Every release of Kaiva Bridge, newest first. Kaiva Bridge is in active development, expect frequent additions.

Latest 16 Oct 2026
  1. 16 Oct 2026Custom domains (Pro)

    Use your own domain with Kaiva Bridge

    Put your MCP servers on your own domain, like mcp.yourcompany.com, with your name and logo. Your customers can sign in with the account they already have with you, no Kaiva account needed. On Pro.

    • NewCustom domains. Add mcp.yourcompany.com once in Workspace settings, Domains, with one DNS record. Each server gets an address on it, like mcp.yourcompany.com/orders, and its Kaiva address keeps working.
    • NewCustomer sign-in. Switch it on in a server's Sign-in tab, and your customers sign in with their account at your API.
    • NewYour brand. Your name and logo on the sign-in screen, plus a setup page at your domain listing your servers, with steps for Claude, ChatGPT and Cursor.
    • NewAlso: a list of your customers with their usage and an optional monthly limit, and a listing kit for your docs and MCP directories.
  2. 15 Oct 2026Verified writes (writes are off by default)

    When your AI changes your data, Bridge now confirms the change and lets you undo it

    Tools that change data are off by default, so your AI cannot change anything until you switch a tool on. For tools you switch on, Bridge can now read each change back to confirm it worked and tell your AI, and a wrong change can be undone within 24 hours. You set this up per tool. This works on REST API and remote MCP servers.

    • NewConfirm each change. Set this up per tool, or let Bridge suggest it from your API's spec.
    • NewUndo. Your AI, or you from the server's Traffic tab, can reverse a change once within 24 hours.
    • NewA Writes list on the Traffic tab shows recent changes and whether each was confirmed.
  3. 14 Oct 2026Free 10 servers

    Free now includes 10 servers, and Pro is unlimited

    • ChangeFree plan raised to 10 MCP servers, up from 3.
    • ChangePro plan servers are now unlimited, up from 15.
  4. 14 Oct 2026CLI 0.1.17

    Set up Kaiva Bridge from your coding agent, and add it to VS Code

    Tell your agent: set up Kaiva Bridge: read https://kaiv.ai/bridge/SKILL.md and follow it. It turns your API, database or documents into an MCP server and connects it to your AI client; a public API needs no account. From kaiv.ai/bridge, one button adds Kaiva Bridge to Claude, Cursor, VS Code, Claude Code or ChatGPT.

    • NewVS Code signs in to Kaiva Bridge directly, like Claude and ChatGPT.
    • Newkaiva-bridge push <name> --db turns the database in your project's DATABASE_URL into a server in one command: Postgres, MySQL, MariaDB or SQL Server.
    • NewEvery push and try ends with a setup check: one real tool call, which tools can change data, and how much of your agent's context the tool list takes.
    • Newkaiv.ai/bridge/auth.md: the no-account recipe for agents, from create to claim.
    • NewRun Kaiva Bridge with NVIDIA's NeMo Agent Toolkit on NVIDIA Brev in one click, from the example in the CLI repository.
  5. 14 Oct 2026NemoClaw

    Kaiva Bridge works with NVIDIA NemoClaw

    NemoClaw, NVIDIA's stack for OpenClaw agents, accepts MCP servers as authenticated HTTPS endpoints, which is what every Kaiva Bridge server is. A server's Connect tab now has a NemoClaw tab with the one command that adds it to a NemoClaw sandbox, and the console names OpenClaw when it connects.

    • NewNemoClaw tab on Connect and in Setup: one command adds your server, key included, and NemoClaw's own check confirms it.
    • ChangeNew server addresses are accepted by NemoClaw as they are.
  6. 13 Oct 2026SQL Server + GraphQL

    Kaiva Bridge now supports SQL Server and GraphQL

    • NewSQL Server databases as MCP servers: connect SQL Server or Azure SQL, and each table becomes read-only tools.
    • NewGraphQL APIs as MCP servers: point Bridge at a GraphQL endpoint and its queries become tools.
  7. 9 Oct 2026WebMCP support (beta)

    Your MCP server on your own website, with WebMCP support (beta)

    AI agents in the browser can now use your server's tools on your own website. Choose which tools are public, list your sites, and paste one block of code into your page. In browsers with WebMCP (Chrome so far), your page offers those tools to the browser's AI agent, and each call goes through Kaiva Bridge like any other. To get started, open a server's Connect tab and choose Your website (WebMCP).

    • NewWebsite keys, made for a public page: they reach only the tools you make public.
    • NewChange a website key's sites and tools without changing the code on your site.
    • NewSee the last tool call from your website on the Connect tab.
  8. 8 Oct 2026CLI 0.1.11

    Your AI agent can create an MCP server before you sign up

    Coding agents such as Claude Code, Codex and Cursor can turn a public OpenAPI spec into a working MCP server with no account, then give you a link to keep it. Run npx @kaiva/bridge try --openapi <spec-url>, or install the Kaiva Bridge plugin for Claude Code.

    • Newkaiva-bridge try: a server with the spec's GET operations for 72 hours, a key for it, and a claim link.
    • NewClaim page: see what the agent made before you sign up, then keep it in your workspace.
    • NewClaude Code plugin: claude plugin marketplace add ai-kaiva/bridge, then claude plugin install kaiva-bridge@kaiva.
  9. 8 Oct 2026cURL + Postman

    Import an API from a cURL command or a Postman collection

    No OpenAPI spec? Paste a cURL command or upload a Postman collection, and Kaiva Bridge turns it into an MCP server. It is read in your browser, and you review every request before anything is created. To get started, create a new server, choose REST API, then cURL or Postman.

    • NewPaste a cURL command, as copied from API docs, a terminal or your browser, and it becomes a tool for that request.
    • NewUpload a Postman Collection (v2.0 or v2.1) and choose which requests become tools. Anything Bridge can't import is listed with the reason.
    • FixBug fixes and reliability improvements
  10. 6 Oct 2026CLI 0.1.10

    Sign in from the command line and add servers to Claude Code or Cursor

    Run kaiva-bridge login and approve the code in the console: no keys to create or copy. Then add a server to Claude Code or Cursor with one command, and check a connection end to end with kaiva-bridge doctor. Update with npm install -g @kaiva/bridge.

    • Newkaiva-bridge login, whoami and logout. Sign in by approving a code in the console instead of copying a key. KAIVA_BRIDGE_TOKEN still works for CI.
    • Newkaiva-bridge install adds a live server to Claude Code or Cursor and keeps your other entries. You can preview it or undo it. Other AI clients get their setup steps.
    • Newkaiva-bridge doctor checks sign-in, publication, the MCP connection and the tool list, and says what to fix.
    • NewStarter sets. Pick the few tools one task needs and give them clearer descriptions, saved as versions you can restore.
    • FixBug fixes and reliability improvements
  11. 2–6 Oct 2026New console

    A clearer Kaiva Bridge console

    A new look across the console and a simpler path from your first server to your first tool call.

    • NewUI and UX upgrades with an improved setup journey, from creating a server to your AI client's first tool call
    • ChangeBetter layouts on phones and smaller screens
    • ChangeClearer, fewer emails
    • FixBug fixes and reliability improvements
  12. 2 Oct 2026MCP 2026-07-28

    Support for the latest MCP specification

    Kaiva Bridge now follows the newest version of the Model Context Protocol, released 28 July 2026, and still answers clients on earlier versions. Nothing changes for your servers or your AI clients.

    • NewFull support for MCP 2026-07-28
    • FixClearer errors when a client sends an unsupported request
  13. 1 Oct 2026Documents

    Documents as MCP servers

    Drop in PDFs, Word files, images or notes and Kaiva Bridge turns them into a hosted MCP server your AI can search. It can read any page, tables stay as tables, and every answer says which document and page it came from. To get started, create a new server and choose Documents.

    • NewScans and screenshots are read as well as typed documents
    • NewSave a search and it becomes a tool of its own
    • NewUpload a new version and it replaces the old one once it is ready, with search working throughout
    • ChangeOne New server screen for every source, and filters by source on the Servers list
    • ChangeRename any server by clicking its name
    • FixBug fixes and reliability improvements
  14. 29–30 Sep 2026MySQL + MariaDB

    Kaiva Bridge now supports MySQL and MariaDB

    You can now connect MySQL and MariaDB databases to Kaiva Bridge. Your tables become read-only tools your agents can call, the same way Postgres works today. To get started, create a new server and choose Database.

    • NewChoose your database, then paste a connection string or fill in the host and login
    • NewA ready-to-copy read-only login for each database
    • ChangeClearer messages when a database can't be connected, including Postgres
    • ChangeA simpler form for issuing API keys
    • FixBug fixes and reliability improvements
  15. 14–15 Sep 2026Spec upload

    Paste or upload your OpenAPI spec, no public URL needed

    OpenAPI specs no longer need to be hosted anywhere. Paste one into the console or upload the file, in JSON or YAML, and Bridge checks it in full before a server is created: a broken spec points to the exact line and column, and one with nowhere to send requests is caught up front. The Connect tab now walks you from published to your first successful tool call, and new database sources verify their server certificate by default.

    • NewPaste or upload an OpenAPI or Swagger 2.0 spec, JSON or YAML up to 900 KB, instead of hosting it at a URL
    • NewAn API base URL field for specs that don't declare a server
    • NewReplace a pasted spec from the server page; the Management API can create and replace one too
    • NewSpec errors point to the exact line and column, and a spec that can't work is refused before a server exists
    • NewConnect tab setup progress: published, client connected, first tool call, updating on its own
    • NewThe server header shows when a client last called a tool, and whether that call failed
    • NewClaude Code and Codex connection tabs; a key created on the Connect tab is filled into every snippet
    • NewNew database sources verify the server certificate by default, with your own CA certificate supported
    • ChangePath-level parameters in a spec are now included in the tools it generates
    • ChangeClearer setup errors when a source refuses the connection or its certificate can't be verified
    • ChangeCreate form: clearer headings and spacing, and the page stays in place when a check fails
    • ChangeImproved platform security hardening
    • FixConsole display and loading fixes
    • FixAssorted fixes across connection and setup flows
  16. 5 Sep 2026Tool integrity

    A rewritten tool description waits for you instead of reaching your agents

    An MCP client reads a server's tool definitions once, when someone approves the connection, and generally never shows them again. What the gateway serves is now a pinned revision rather than whatever the last introspection wrote, so a changed description or input schema on a tool that is already exposed waits for you to publish it. Connected agents keep receiving the definitions they had. New and removed tools are unaffected and still publish with no interaction. Every live server also publishes the fingerprint of what it serves, and the tool definitions behind it, so anyone can check without an account.

    • NewChanged descriptions and schemas on exposed tools wait for a decision, with the tools named
    • NewPublish or reject from the server page; rejecting restores what is being served
    • NewAuto-publish per server, for workspaces that introspect from CI
    • NewPublic /fingerprint, /contract and /revisions on every live server, no key needed
    • NewEmail and a webhook when a change is waiting
    • NewManagement API: list revisions, promote, reject and set auto-publish
    • NewCLI 0.1.5: introspect, revisions, promote, reject and auto-publish; introspect exits 3 when a change is held
  17. 31 Aug 2026Console

    Per-server views, live tool calls that link through, and errors that explain themselves

    Servers opens as a list you scan and a server you work on, each with its own audit log. Live traffic entries link straight to the exact call in Tool calls, an expanded row shows which entry it belongs to, and a failed call reports the reason rather than only that it failed.

    • NewPer-server view, with the audit log scoped to that server
    • NewLive traffic entries open the exact matching call in Tool calls
    • NewFailed calls report what went wrong
    • ChangeExpanded audit rows show which entry they belong to
    • ChangeUI pass across the console
  18. 29–30 Aug 2026OAuth 2.1

    One policy model, whatever the client

    Per-tool policies and require-approval apply to every caller, whether it authenticates with an API key or over OAuth 2.1. A policy governs a tool by the tool (not by how the client connected) and every call lands in the audit log the same way.

    • NewPer-tool policies across API-key and OAuth 2.1 clients
    • NewRequire-approval across API-key and OAuth 2.1 clients
    • NewOAuth 2.1 calls in the audit log
    • ChangePolicy subjects can target OAuth clients directly
  19. 28 Aug 2026Free tier 4×

    200,000 tool calls a month on Free, and what your tools cost the agent

    The free tier is four times what it was: 200,000 tool calls a month, still with no card and no expiry. Alongside it, the console now shows what each tool definition costs your agent in context, and which tools it has never called, so you can switch them off.

    • ChangeFree tier raised to 200,000 tool calls / mo, up from 50,000
    • NewContext cost per tool definition, shown in the console
    • NewUnused tools flagged so you can unexpose them
  20. 27 Aug 2026Bundles

    One endpoint for many servers

    A bundle collapses several servers into a single MCP endpoint. tools/list merges across its members and tools/call routes to the server that owns the tool, so a client config has one entry and one key instead of one per server, and adding a server later needs no change to the client.

    • NewBundles: one endpoint and one key across several servers
    • NewMerged tool list, with calls routed to the owning server
    • NewKeys are scoped to a bundle, re-checked on every call, so a server added later is not reachable by an older key
    • ChangeColliding tool names are prefixed on both sides, so a name never changes meaning as membership changes
    • ChangeOverview shows trends and sparklines, not four bare numbers
  21. 27 Aug 2026Teams + SSO

    Teams on every plan, and single sign-on expanded

    Free workspaces now include three members instead of one, so a small team can share a Bridge workspace without paying for it. Single sign-on now covers SAML 2.0 alongside OIDC, with email-domain verification and provider groups mapped to Bridge roles.

    • NewFree plan now includes 3 team members, up from 1
    • NewTeam invitations, invite by email with a role, expiring after 7 days and cancellable at any time
    • NewSingle sign-on now covers SAML 2.0 alongside OIDC, Okta, Entra ID, Google Workspace or any compliant provider
    • NewVerify your email domain and members sign in with their work email
    • NewProvider groups can now map directly to Bridge roles
  22. 24 Aug 2026CLI 0.1.4

    Upstream compatibility

    Some APIs refuse requests that arrive without a User-Agent. Bridge now identifies itself on every upstream call, which fixed introspection and tool execution against those hosts.

    • FixSend a User-Agent on all upstream requests, GitHub and others returned 403 without one
    • NewListed on Glama
  23. 22–23 Aug 2026MCP 2026-07-28

    Dual-era MCP, and Bridge as a connector

    Full support for the 2026-07-28 MCP revision alongside the legacy handshake, so new and old clients both work against the same endpoint with no flag to set.

    • NewBridge itself is an MCP connector, create and inspect servers from your agent
    • NewTool annotations derived from the HTTP verb, so clients know what is read-only
    • NewoutputSchema generated from documented OpenAPI responses
    • ChangeDual-era protocol negotiation; legacy clients untouched
    • FixClosed a dynamic-client-registration bypass
  24. 20–21 Aug 2026CLI 0.1.0–0.1.2

    Management API and a CLI on npm

    Everything the console does is now an API, and the CLI is public on npm as @kaiva/bridge, MIT, source on GitHub.

    • NewManagement API with scoped management keys
    • New@kaiva/bridge published to npm: tools command, invoke a tool by name
    • Newemail-check source: MX, SPF and DMARC lookups with a verdict
    • ChangeRe-introspection is non-destructive and keeps OpenAPI descriptions and constraints
  25. 19–20 Aug 2026Registry

    Remote MCP sources, playground, and a registry listing

    A third source type: point Bridge at an existing remote MCP server and govern it the same way. Listed in the official MCP registry as ai.kaiv/bridge.

    • NewMCP proxy source, bring a remote MCP server under Bridge
    • NewIn-console playground, shareable server links, and templates
    • New/try sandbox and a live runner on the landing page
    • NewListed in the official MCP registry as ai.kaiv/bridge
  26. 17–19 Aug 2026OAuth 2.1

    OAuth 2.1, end to end

    Full authorization-code flow on Ory Hydra with dynamic client registration, so an MCP client can connect without anyone pasting a key.

    • NewOAuth 2.1 authorization-code flow with PKCE
    • NewDynamic client registration, rate limited
    • NewRFC 9728 / RFC 8414 discovery at /.well-known, plus security.txt
    • NewConsent screen with human-readable scope labels
    • NewConnected apps: view and revoke authorized clients yourself
    • FixA token issued for one server can no longer be replayed against another
  27. 15–16 Aug 2026Demos

    Try it without an account

    An isolated public demo host, so anyone can call a real MCP tool before signing up for anything.

    • NewKeyless public demos on an isolated host
    • NewEdge-only ingress, the origin refuses direct traffic
    • NewAbuse auto-suspend, on by default
    • NewIn-console bug reports and feedback
    • NewSign in with Google
    • FixOpenAPI connector: servers[] base URLs and $ref parameters
  28. 12–14 Aug 2026Plans

    Plans, status, and isolation at the database layer

    Row-Level Security so one workspace cannot see another's servers, keys or logs, enforced by Postgres, not by application code.

    • NewRow-Level Security and org isolation
    • NewFree / Pro / Enterprise plans with Stripe
    • NewPublic status page with 90-day uptime, independently monitored
    • NewPer-client connect tabs and current MCP protocol negotiation
    • ChangeConsole and site work properly on a phone
  29. 9–11 Aug 2026Hardening

    Egress guard, RBAC, and a kill-switch

    Nothing reaches a private network, nothing runs without a scoped key, and there is one switch that stops everything.

    • NewSSRF egress guard, private ranges and cloud metadata addresses blocked
    • NewCSRF protection, RBAC, and a global kill-switch
    • NewTOTP two-factor on admin
    • NewDPA, plus Bridge sections in Terms and Privacy
  30. 8 Aug 2026v1.0

    Kaiva Bridge, first build

    Turn an OpenAPI spec or a Postgres database into a live MCP server. One endpoint, one key, upstream credentials kept server side.

    • NewSources: OpenAPI documents and Postgres databases
    • NewLive MCP gateway with scoped API keys
    • NewPolicies (allow, hold for review, or deny) with dry-run against recent traffic
    • NewHuman-in-the-loop approvals with webhook delivery
    • NewAudit log with tamper-evident verification and export
    • NewResponse DLP, per-key rate limits, and usage enforcement
    • NewSCIM provisioning and SIEM streaming for teams