Changelog

What changed in MadMax, week by week. New blocks and integrations, improvements to the canvas and the AI builder, and fixes, written for the people who use it.

Week of October 26, 2026

The free plan runs 20 active workflows, the docs answer questions, and the site loads faster, measured at 100 on PageSpeed this week on desktop and mobile.

New

  • Ask the docs. Every /docs page has a panel that answers questions from the documentation itself, with links to the pages it drew from. No account needed, and it draws on the public documentation only, not on your workflows or connections.
  • Notes on workflows and folders. The list shows one icon per item: i for notes, or an amber warning mark when something still needs setup (a blank connection, a required field, a missing store, a dangling workflow link). The modal lists what is missing with a "fix with Max" link that opens the canvas with the prompt ready. Notes render as markdown, and folder notes can be set over the REST API.
  • Turn a workflow on or off from its notes modal without opening the canvas, or switch a whole folder at once. The switch waits until setup is finished.
  • Pages for every integration at /integrations and for alternatives to Zapier, Make, n8n, Tines, Workato and Pipedream at /alternatives.

Improved

  • The free plan runs 20 active workflows, up from 3, and the cap counts active workflows at activation. Creating, importing and installing templates is not capped; activation tells you when the limit is reached.
  • The site loads faster. madmax.build was measured at 100 on PageSpeed this week for performance, accessibility, best practices and SEO on desktop and mobile.
  • The docs grew. Forty-one pages were expanded and checked against the code, every step page now covers each Inspector field and the exact output shape, every page has a search-friendly title and description, and there is a new page for the workflows index.
  • Using a template shows it is installing, and a failed install says why on the page.

Fixed

  • Installing a large bundle with many Run code steps, such as the GitHub PR Review Bot, is fast and reliable, and every Run code step runs faster.
  • The canvas keeps the saved flow direction when a workflow or an installed bundle opens, so ports sit on the expected sides of each node.
  • www.madmax.build redirects to madmax.build.

Week of October 19, 2026

Removing a connection cleans up after itself, notification emails share one design, and the landing page shows real workflows.

New

  • Platform emails have one design: run failed, connection needs reconnecting, approval needed, workspace invite and sign-in code, each with the facts you need (the failed step, timing, the error, who invited you). Two new receipts: an API token was minted, and a workspace moved to Pro.
  • The landing page's "How it works" shows four example workflows on the real canvas, with the prompt, the steps Max built, and a sample run log.
  • Signing in asks you to accept the current terms of service once. The terms gain sections on suspension and on how changes are announced.

Improved

  • Connection cards in Settings edit the connection's own settings, such as a default channel for Slack, instead of a raw JSON box. When Max wires that connection into a step and leaves the channel blank, the default fills in.
  • Removing a connection tells you first which workflows use it and how many are live, then clears the steps that used it and turns those workflows off until a new connection is chosen. Steps that reference a connection removed earlier are cleaned up the same way.
  • Phone numbers used for text messaging are encrypted at rest.
  • The template store is a curated catalog: publishing is by MadMax only, and the tabs, category filters and fork option are gone. Search, sort, app and trigger filters stay.
  • The live power button is a toggle again. Turning a workflow on keeps you on the canvas, edits publish through the arrow button, and clicking the live button turns it off.

Fixed

  • Improved verification of branch conditions before activation: a condition that compares a trigger field against an example value from a documented sample is flagged, and generated test samples put ids where ids belong, so branches evaluate the same way on real events.
  • The plan no longer shows a verifying badge for a step Max removed, or a "resume the build" button with nothing to resume. Steps Max deletes retire their plan rows.
  • "Use the block builder" on an empty canvas dismisses the scaffold card again.
  • The left rail's collapse button stays in view when the Chat, Blocks and Plan tabs are wide.

Week of October 12, 2026

Max runs its own verification loop and can look things up on the web; Slack gains reaction triggers and a public install page.

New

  • Slack reactions are a trigger. A workflow listening to Slack app events now receives emoji reactions alongside messages and @-mentions, pointed at the message that was reacted to so a step can reply in its thread.
  • Max can look outside the platform while building: search the web, read a page or PDF, pull up a runbook, fetch the exact spec for an integration operation, and research a third-party webhook's documented payload so the shape it builds against is the real one.
  • Verification is available over MCP. External clients can call verify_workflow_run to get the same report the canvas uses, repair, and verify again.
  • Public pages for the Slack app: an install and overview page at /slack, a support page at /support, and the list of sub-processors at /sub-processors.

Improved

  • One build loop. Max proposes the plan and builds in the same turn, then runs verification once, repairs what it names, and verifies again until every step passes or a human is needed. It stops only for a connection, a real sample, or something that does not exist. The chat offers Build and Ask; Edit folded into Build.
  • Slack and GitHub event triggers, and forms, verify on their documented shape. Max no longer asks you to @-mention the bot or submit the form before it can continue; a real event still replaces the sample when it lands.
  • When a provider has one shared connection and one of your own, Max picks yours instead of asking. Branches that only a rejected approval or a different event would reach now get their own verification pass.
  • Settings has a "← workflows" link back to the list.

Fixed

  • Build turns in the canvas chat run through every planned step to completion in a single turn.
  • Steps after an integration step receive the same data shape in verification and in live runs (payload.channel, for example), so what verifies is what runs.
  • Improved workflow verification to catch data references the upstream step cannot satisfy before activation, and a real captured sample re-checks the steps downstream of it.
  • Live deliveries run the active version of a workflow, including while the canvas is waiting for a test sample.
  • Connection validation checks the provider: a step only accepts a connection for its own integration, so a Slack step cannot be pointed at a Gmail connection.

Week of October 5, 2026

Secrets for any step, encrypted fields in data stores, and runs you can start with data from the REST API or MCP.

New

  • Keep API keys and tokens in one Secrets connection and reference them from any step as {{ secrets.NAME }}. Values are stored encrypted and masked in run output and pinned samples, and a missing name fails the step with an error that names it. Manage them in Settings under Connectors, or ask Max to store one.
  • Mark a data store field as sensitive and its value is encrypted at rest inside each row. This is encryption at rest, not access control: steps, exports and authenticated reads over the REST API return the value, while the Settings browser masks it.
  • Start a workflow with data. POST /workflows/:id/run and the MCP run_workflow tool take an inputs object, which lands as the trigger's payload the same way a Run from the canvas does.
  • The template store is open again from the account menu, with the official Company Slack Assistant and GitHub PR Review Bot templates.

Improved

  • Settings has a redesigned Developer tab: mint a token first, then pick how to connect a client (MCP client, REST, or the agent brief) with ready-to-paste snippets. Tokens page ten at a time.
  • Deleting a data store or one of its rows is now limited to workspace owners and admins. Members keep reading, writing and editing the schema.
  • Runs with thousands of steps open fast. The run page loads each step's input, output and logs only when you open its card.
  • A folder created, renamed or deleted from the REST API, MCP or another tab shows up on the workflows list without a reload, and an open canvas picks up a workflow turned on or off elsewhere.
  • Improved deployment handling so running workflows can complete within their execution limits before a new release takes over.

Fixed

  • Reconnecting a GitHub connection that already had the MadMax app installed now completes instead of stopping on GitHub's "already installed" page. If your GitHub authorization expires, workflows keep running through the installed app and you are asked to reconnect.
  • {{ secrets.NAME }} resolves when a step is tested from the Inspector or verified by Max, the same as in live runs, so an authenticated request is tested authenticated. Previews show a placeholder in place of the value.

Week of September 28, 2026

Webhook auth over the REST API, exports that carry their data stores and memories, a token budget for agent answers, and stronger isolation between runs.

New

  • A webhook trigger's inbound auth can be set over the REST API: read the trigger URL and auth method in force, switch between bearer and HMAC (GitHub, Stripe or plain format), and have a secret minted for you. The secret is returned once, on the call that set it, exactly as the canvas shows it.
  • Exports are portable. A workflow export carries the definitions of the data stores and AI memories it uses and refers to them by name, and importing into another workspace creates or reuses them before the workflow lands. Rows and secrets are not exported.
  • The AI Agent step has a max_tokens setting. Single-shot calls default to a 16,000 token output budget instead of the provider's 4,096.

Improved

  • Improved reliability of large knowledge base backfills by pacing embedding requests against provider limits.
  • Each run has a memory budget, with a plain message naming the limit when a run reaches it. Improved resource isolation and workload stability.
  • The GitHub PR review bot template responds to mentions in edited comments, acknowledges with reactions, checks that inline comments anchor to real diff lines, and falls back to a regular comment when GitHub rejects an inline review.

Fixed

  • An AI Agent answer cut off by the output limit fails its own step with "output truncated" and the remedy, so the problem is reported where it happens.

Week of September 21, 2026

Workflows can write to knowledge bases, Gitea joins the integrations, MCP clients manage data stores and activation, and loops run eight iterations at a time.

New

  • Save to knowledge, a step that writes a document into a knowledge base (or deletes one) from inside a workflow. Source code is chunked with awareness of its structure, knowledge bases can be referenced by name as well as id, and the AI Agent can be told what to look up with a focused knowledge query instead of its whole prompt.
  • Gitea, self-hosted. Point a connection at your instance, sign in with OAuth or paste an access token, and use the same repository, pull request and comment operations the GitHub integration offers, plus push, pull request and comment triggers.
  • MCP clients can list and create data stores, and can activate or deactivate a workflow once you confirm in your client's chat.
  • Loops run up to eight iterations at a time. Results keep the list's order and failures are counted per item, so large loops finish in a fraction of the time.
  • The GitHub PR review bot template keeps a knowledge base of your repository, refreshed on every push to the default branch, and the reviewer reads from it while reviewing. Exports now carry knowledge base references by name and imports create any that are missing.

Improved

  • Every import lays out the canvas: template store Use and Fork, n8n import, Duplicate and REST imports place steps that arrive without positions instead of stacking them at the origin.
  • Folder export shows progress. The button disables while the file builds and the download starts when it is ready, so a slow export no longer looks like nothing happened.
  • Run history keeps your place: an expanded loop stays open while the timeline refreshes. The workflows list updates on its own when a workflow is created, moved, archived or activated from the API, MCP or another tab.
  • Run code has improved performance with larger inputs, and an error you throw reports its own message so you can see exactly what went wrong.

Fixed

  • Step cards on the canvas show map and list values from a step's configuration (owner: ... · repo: ...) instead of "[object Object]".
  • Bulk saves into a brand-new knowledge base land every document, including when the embedding provider is busy.
  • The GitHub GraphQL operation puts the response under result, so a workflow reads payload.data.result instead of a confusing data.data.
  • A run retried mid-loop resumes at the exact iteration it stopped on, and the canvas loads only the top-level steps of recent runs so long loops do not weigh it down.
  • Send text steps whose phone number was removed fall back or report a clear error, inbound text menus stay unambiguous when two entries share a label, and fetching an inbound email is more resilient to transient provider errors.

Week of September 14, 2026

Whole folders as template store bundles, folder export and import, OpenRouter and a browser tool for the AI Agent, a new Switch editor, and fewer dead ends for MCP clients.

New

  • A whole folder of workflows can be published to the template store as one bundle. Use or Fork installs the folder with every member workflow, and the store shows how many workflows a bundle contains.
  • Folders export from the folder menu, with in-folder call references kept intact. The Import modal accepts a folder export or a .zip: a zip of single exports becomes a new folder named after the file.
  • OpenRouter is available as an AI Agent connector. Bring your own key and paste any OpenRouter model id, routing suffix included; leave the model blank to use the platform default.
  • The AI Agent step has a built-in browser tool: it can read a web page or PDF, or open an interactive session to click and type through a site, with safeguards on page size and session length.
  • Switch has a new case editor: ordered cases in their own modal, each evaluated as you type against the pinned sample so you can see which branch a real payload would take before you save.
  • GitHub connects through the MadMax GitHub App, and the GitHub integration gains operations to create and complete check runs, react to comments, and list, update or delete reviews and review comments.

Improved

  • The Inspector describes the option you picked: the hint under a segmented control (a webhook's response mode, a call's wait or detach mode) changes with the selection. Multiline Switch expressions and Set fields values grow as you type, and every step shows its stable id for use in references.
  • The live overlay folds long step lists behind a "more steps" toggle and scrolls as a panel, so the checklist, the 24 hour stats and the run buttons stay reachable on a 30-step workflow.
  • Build mode talks first. "Hello" or a question gets a normal reply that invites a description; Max only proposes a plan when your message describes an automation, and never on a canvas that already has steps.
  • Improved reliability of scheduled workflows and background jobs across deploys.
  • Improved credit metering precision for smaller usage increments.

Fixed

  • Max wires every branch of a Switch, If or Loop that feeds the same step, so each branch reaches the merge.
  • Validation of AI Agent steps built over MCP covers nested objects in the output schema, so problems surface before a run, and Extract agents with their own fields validate and clear their findings cleanly.
  • External MCP clients get further on their own: building after a proposed plan counts as approval, a webhook test delivery pins the same shape a real delivery would and can start a draft run, and running a webhook-started draft reports the actual outcome. Schedules gain the common US and European time zones.

Week of September 7, 2026

Sub-workflows that run in parallel, merge writes to data stores, defaults for optional fields, a map of how workflows call each other, and safeguards that keep one workflow from affecting others.

New

  • Call workflow has a "detach" mode. The step starts the sub-workflow and moves on right away, so one listener can hand work to several sub-workflows at once, each with its own run and its own approvals. The default "wait" mode is unchanged.
  • Save to store has a "merge" mode. Set two fields and the other ten are kept: the fields you set merge into the existing row, nested maps merge, and anything you did not touch survives. "Replace" stays the default.
  • Templates accept a default for fields that may be missing: {{ payload.x | default:"" }}, | default:0 or | default:false. Optional fields no longer need a Run code step just to fill a blank, and verification treats a defaulted field as resolved.
  • Max and MCP clients can rewire a workflow in one pass: remove, add, update, disconnect and connect steps in a single edit, validated once at the end instead of after every partial change. Disconnecting a connection that is already gone is reported as such and the rest of the edit goes ahead.
  • The workflows index has a map view showing which workflows call which sub-workflows. Workflows also carry a description (shown on cards, and on hover in the table), and folders can be dragged to reorder them, contents included.
  • The REST API can write data store rows (PUT and DELETE on a store's rows), create a workflow from a document, and auto-lays out a graph you PUT so steps do not pile up at the origin. The API reference gains a Folders section.
  • MCP clients can discover the REST API: the OpenAPI spec and the REST guides are available as MCP resources, and Settings → developer has a copyable "Hand this to your agent" brief. A new "MCP or REST?" page in the docs explains when to use which, and Docs is now one click away in the account menu.
  • Gmail gains a rename label operation. The label keeps its id, so every message already carrying it keeps it under the new name.

Improved

  • Runs are fenced. Every run operates inside safeguards for memory, time and payload size, sub-workflow calls are guarded against runaway recursion, schedules have a minimum interval, and webhook triggers are protected against bursts, reducing the risk of one workflow affecting others. A workspace that has used its monthly run budget is told so in plain language before a run is created.
  • The free plan now includes 500 runs a month (up from 100). The pricing page, the FAQ and Settings → Billing read the same numbers the limits enforce.

Fixed

  • Execution follows the graph exactly: only steps reachable from a trigger run, matching what validation shows on the canvas.
  • == in If, Switch, filter and Loop until expressions now matches a real boolean against "true" and a number against "2", the way values arrive from AI agents and JSON. A x in list check on a missing field now says "expected a list, got nothing" instead of a cryptic error.
  • The REST API accepts the standard JSON:API Accept header.
  • Call workflow in "wait" mode inside a Loop runs each sub-workflow to completion within its iteration, so the loop reports the batch complete only once every sub-workflow has finished.

Week of August 31, 2026

Workflow folders, a REST API for your data, approvals over Slack, email or text, a Google Calendar integration, and a round of Slack and API fixes.

New

  • Folders on the workflows index. Create folders with one level of subfolders, drag workflows between them (a drop strip appears while dragging for moving one back to the top level), rename and delete folders (their workflows drop back to the top level), and keep your own sort order. Folders are on the REST API too, and a folder exports as one JSON document carrying every workflow in it; links between workflows in the folder survive the import.
  • The REST API covers your data: create, update and delete knowledge bases (and add documents to them), AI memories, and data stores. MCP clients and Max can now see your knowledge base ids, memory ids, store slugs and verified phone numbers, so they can fill those fields directly. A failed step's error is included in the execution steps endpoint by default.
  • Approval steps choose a channel: Slack, email or text, with a list of approvers for each. Slack approvers (user ids or emails) get a direct message with Approve and Reject buttons the moment the run pauses, email approvers each get signed one-click links, and verified phone numbers get a text answered with YES or NO.
  • Google Calendar integration: check free/busy, list events, and create, update or delete events. Attendees receive invites.
  • The Email trigger accepts forwarded inboxes: an "any authenticated sender" mode for mailboxes auto-forwarded to the step's address, and the allowed-sender field takes a comma-separated list of addresses.
  • AI Agent output includes tool_names, the list of tools the agent used, so an If step can branch on what the agent did.

Improved

  • The canvas opens faster.
  • The approval decision page you reach from an email link is dark like the rest of the product, and branded emails show the MadMax icon across mail clients.

Fixed

  • Slack bots reply to follow-ups in their own threads. Existing Slack connections pick this up on the next inbound message, no reconnect needed, and the run logs say why a follow-up was skipped.
  • Data stores, knowledge bases and memories can be updated and deleted through the REST API. Exports leave out workspace-specific ids (connections inside AI Agent tools, approval Slack connections, verified phone ids), so they import cleanly elsewhere.
  • Improved reliability of AI Agent model connections: a stalled response fails promptly and the step's retry uses a fresh connection.
  • Posts from other Slack apps and API integrations can trigger workflows; the self-reply guard applies only to your own bot.
  • The Watch trigger finds the list of items inside nested responses (Gmail message search, for example).
  • Approval notifications post to the chosen Slack channel, and the Inspector shows only the fields for the chosen approval channel.

Week of August 24, 2026

Approvals from Slack and the run page, Slack app events as a trigger, more accurate and encrypted knowledge bases, and a docs site that grew by more than fifty pages.

New

  • Approvals can be decided from Slack and from the run page. Point an approval step at a Slack channel and it posts Approve and Reject buttons; once someone clicks, the message rewrites itself to say who decided. Any member of the workspace who can open the run can also decide a waiting approval from its page, so an approval can be settled without a notification email.
  • The Webhook trigger has an App-event source: connect Slack and the workflow fires on Slack events such as mentions and direct messages, with a normalized payload, signature checks, de-duplication, and a guard that stops a bot from replying to itself.
  • Slack gained four operations: add a reaction, look up a user by email, get a user, and read a thread.
  • AI memory compacts long sessions. Once a session outgrows its window, older turns fold into a rolling summary, so an agent keeps the whole conversation's context at a predictable cost. Compaction is on by default and can be turned off per memory.
  • The docs grew by more than fifty pages: a Building with AI section (Build mode and the plan, how verification works, Edit mode, Ask mode, the chat, AI settings), a reference page for every block, Runs (run history, activation and snapshots, testing, error handling, waits and approvals), Data (data stores, AI memory, knowledge bases), an AI workflow automation overview, and comparisons with n8n, Zapier and Tines. A search box sits at the top of the docs sidebar (press / to focus it), and the pager and sidebar links now land at the top of the new page.

Improved

  • Knowledge base retrieval was reworked. It combines keyword and meaning-based search, every chunk gets a one-line context header at upload, a re-ranking pass orders the candidates, and each passage is labeled with its source file so agents can name their sources. Improved retrieval accuracy in internal testing.
  • Knowledge base documents now benefit from encryption at rest.
  • Max picks the right block more reliably: "look up the customer record from our table" goes to Get from store, while a named database (Postgres, Airtable, Sheets) goes to an integration. Integration operation search is broader too, so asks like "upload a file to Google Drive" or "read recent Slack messages" find the right operation.
  • Max spends fewer turns on fixes: it batches validation findings into one pass, and AI Agent steps that call tools keep their reasoning between tool calls.
  • Blocks palette: an AI Agent step with no connection of its own shows "MadMax default" as its model, Custom prompt is the first AI Agent tile, and the Fetch section is gone (Get from store sits under Act).
  • The template store is hidden for now. The imported templates were removed; the store returns when hand-built templates are ready.

Fixed

  • Switch edges with custom case labels show on the canvas and in the routes view, and every case label has its own output port. Adding a Slack reaction that is already present no longer fails a run.
  • The Webhook trigger's App picker stays visible and changeable after an app is chosen. Repeated hints under the Webhook, File upload and Watch triggers were trimmed, the AI Agent step shows a single Advanced section, and opening a block launcher closes the plan peek instead of painting underneath it.
  • Strengthened workspace isolation and data access validation.

Week of August 17, 2026

Build, Edit and Ask modes for the chat, branded emails, a Settings tab for data stores, and tighter plan and data-shape checks for Max and MCP clients.

New

  • The chat has three modes. Build (Max plans, builds and verifies a workflow) and Edit (changes to an existing one) replace the old Guided and Advanced names. Ask is new: a read-only mode that answers questions about the workflow and cannot change anything. Switch to Ask and back and the workflow is exactly where you left it.
  • Emails from the Send email step wear the MadMax design: a dark header with your workspace and workflow name, your markdown body styled for mail clients, an optional button (set a label and a link), and a footer with the run details. The Inspector's preview shows what recipients see.
  • Data stores have their own tab in Settings.
  • MCP access is included in the Pro plan.

Improved

  • The Inspector, the panel where you configure each block, has been reworked.
  • Plans are checked before you approve them. Every step in a proposed plan has to name a real block or integration, so a plan only promises what exists. Asking to "save the results to my knowledge base" now plans a data store write instead (at the time, knowledge bases were filled by upload and read-only from workflows; workflows can write to them since September), and the plan shows that substitution so you see the decision up front.
  • A step cannot pass verification while a required field is still blank.
  • The plan pill in the chat dock shows a green check and N of N once every step is verified, so you can tell a build is done without opening the plan.
  • Validation understands the shape of each block's output (the webhook body, agent data, approval decisions, store reads and more) and catches references to fields that will not exist at run time. MCP clients get the same findings in every staging response, plus a short build contract when they connect.
  • When a test run references a field that is not in the upstream data, the result names the missing references and lists the paths that do exist, so a repair takes one round instead of a guess.

Fixed

  • Improved connection token refresh scheduling and reliability.

Week of August 10, 2026

Live runs now count as verification evidence, data stores export and import, loop verification is more accurate, and AI usage from live runs shows up in usage views and billing.

New

  • Data stores can be exported and imported, so you can move a store's rows out of a workspace and back in.

Improved

  • Live runs count as verification. When an active workflow completes a run, each step's real output becomes evidence in the plan, so a workflow with a history of clean runs shows that in its verification status. Steps inside loop bodies count too.
  • When you change a step after it was verified, the canvas says so: the step's proof is revoked and Max re-verifies it automatically.
  • Max is steadier on long builds. Verified steps are left alone mid-build, and an AI Agent step that failed its check stays failed until the model re-judges it or you change its settings.
  • Web search is a checkbox in the AI Agent step's Tools list, next to calculator, web request and code, instead of a separate toggle. Existing workflows keep working as they were.

Fixed

  • Loops verify against their body's real output, so loop verification reflects the items the loop actually processes.
  • AI usage from steps in live runs appears in usage views and billing alongside test runs. Usage on your own API key is not billed against your credits.

Week of August 3, 2026

Max now builds the whole workflow in one pass and verifies it in one run, data stores arrive, and verification results are more consistent from round to round.

New

  • Max builds the whole workflow in one pass and verifies it in one run. Instead of placing and testing one step at a time, Max stages every step and edge from the approved plan, then the platform runs the entire graph against realistic generated test data, running reads for real and previewing writes. One review turn confirms what passed and repairs what did not. Steps after a trigger that is still waiting for a real sample verify on generated data first and re-verify once the real delivery lands.
  • Data stores land: a place in your workspace to keep records that workflows save to and read from.

Improved

  • When verification needs your input, it stops at a clear question. Every step it could not finish gets an entry in the plan panel's asks card telling you what to check, and replying continues the walk. Between rounds the walk continues on its own.
  • Loop and Loop until steps verify against the real output of their body, Loop until emulates one iteration and says so, and Call workflow steps preview their result during verification.
  • During whole-graph verification, read-only integration steps such as GitHub get file run for real, so the steps after them verify against actual data.
  • A new check flags an edge that runs from inside a loop body straight into the loop's done path, naming the edge and explaining that the done path would run once per iteration. Max corrects it while building.

Fixed

  • Verification results are more consistent across repair rounds: a step keeps its existing proof when a repair leaves its output unchanged.
  • Steps on a branch the real sample does not take are verified against a generated variant that exercises that branch.
  • Improved Max's responsiveness when several builds run at the same time.

Week of July 27, 2026

Connections become private or shared with workspace roles behind them, Max asks which account to use when several are available, and workspace billing arrives.

New

  • Connections now belong to the person who added them and are private by default. Shared connections are available to everyone in the workspace; a private one is yours alone, and Max cannot see or bind it when someone else is building. Flip the badge in Settings to share or claim one.
  • Workspace roles gate the destructive actions: only the creator or a workspace admin can delete a workflow or edit and remove a connection, and the last owner of a workspace cannot be demoted or removed.
  • Workspace billing: a Pro plan and AI credit top-ups, paid by the workspace owner through a hosted checkout. AI usage meters against the workspace's credit.

Improved

  • The Operation picker in the Inspector was redesigned. The operation name is the control, the description sits underneath, there is search above six operations and full keyboard navigation, and changing the operation clears the old arguments.
  • Improved connection selection when multiple accounts are available. When a provider has two or more connections, a card on the canvas lists the candidates and you choose; Watch steps follow the same rule.
  • When a verification fails because a connection is missing or a service is unreachable, Max hands it to you instead of trying to repair the step. Once you connect the account, MadMax wires it into every waiting step itself and verification resumes.
  • Proposed plans are checked for duplicate step intents, so each step appears once.

Fixed

  • Run step in the Inspector shows its progress: the button spins and the step on the canvas animates while a test runs.
  • AI usage is billed on the provider's actual charge for each call rather than an estimate.

Week of July 20, 2026

AI Agent steps run on MadMax's AI with no key to add, GitHub gains eight operations, and the Inspector shows field help and clearer test failure messages.

New

  • AI Agent steps run on MadMax's AI by default. You no longer need to connect your own AI key before an agent step works; the Inspector shows "MadMax default" as a real selection, and binding your own key stays an explicit choice.
  • Eight new GitHub operations: list tags, get branch head, create branch, get file, put file, merge pull request, delete branch and update issue. Enough to watch a repository and open a pull request from a workflow.
  • A Stop button in the chat halts Max mid-turn.

Improved

  • Max and platform AI steps run on a new default model, picked after a side-by-side bake-off on real builds.
  • Plan rows now name the kind of step they will become, for example "Catch the invoice email (webhook)", so you can see the machinery before you approve.
  • When a workspace has several connections for the same provider, Max wires one and tells you how many alternatives exist so you can swap it in the Inspector. Switch and If steps with a configured case but no outgoing edge are now flagged while building.
  • When a verification fails, Max repairs the step itself and re-verifies instead of stopping to ask. Loops now verify their downstream steps against a real item, and Merge steps verify with the same shape they produce at run time.
  • Every field in the Inspector shows its help text underneath, the Watch block shows real Owner and Repository fields instead of a raw blob, and the Test tab explains what each panel shows and what verification runs against.
  • Gmail's create label operation now reuses an existing label with the same name, and a list labels operation was added.

Fixed

  • Run step in the Inspector now shows the specific reason a test failed, and when several arguments are missing it names all of them.
  • The canvas chat is more reliable: idle turns end cleanly with a plain message, connection errors read as sentences instead of raw error text, and a turn interrupted before it produced anything is retried automatically.
  • The Run code block ships version comparison helpers for semantic version checks, and a run that fails on a missing reference now stops with a clear error instead of retrying.
  • The run detail page updates live as steps finish, and a step's run history shows only real runs; the test note appears only when there are no real runs yet.

Week of July 13, 2026

Email and text messaging arrive as blocks with no setup, the public docs, privacy and terms pages go live, and the chat moves to a left dock beside the Inspector.

New

  • Send email from any workflow with the new Send email block. It sends from MadMax's own address, so there is nothing to connect; the body takes markdown, and while you test it previews the exact email instead of sending.
  • Approval steps can email the approver. The message carries the prompt and one-click Approve and Reject links that work for seven days, and a second click lands on a calm "already decided" page.
  • Start a workflow from an email. The Email trigger block mints its own address; only mail from the sender you allow (your login email by default) gets through, and attachments arrive ready for the extraction steps. Press run, email the address, and that delivery becomes the trigger's sample.
  • Start a workflow from a text. The Text trigger block listens on MadMax's number for texts from numbers you have verified in Settings, under messaging. A personal workspace holds one number, a team workspace holds ten. If a text could run several workflows, MadMax replies with a numbered list and runs the one you pick.
  • The product docs are public, with new pages for Email and Text messaging, and MadMax now has published privacy and terms pages.

Improved

  • The chat docks on the left as a 500px column instead of covering the Inspector, so a step's settings stay open while Max talks. Cmd+I or Ctrl+I opens it, and a Plan or Direct pill lets you choose how Max builds.
  • Every step in a proposed plan now shows its data contract, what goes in and what comes out, so you approve the shape of the data as well as the intent. The plan panel itself was redesigned.
  • The Workflows index was refreshed: a health strip for active workflows, runs and failures over seven days and anything needing attention, a richer table with node counts, last run, edited date and author, and a card menu with rename.
  • Gmail's get message operation now returns the subject, sender, recipients, date and message id, so replies thread correctly and classifying workflows see more than a snippet.
  • The "add a connection" dialog on the canvas shows each provider's logo instead of a letter tile.
  • The homepage was redesigned with use cases and the email and text messaging story.
  • The Watch block gets pickers for its settings in the Inspector, and the AI Agent block's configuration got a usability pass.

Fixed

  • Gmail trash, untrash and label operations configured in the Inspector now run as expected.
  • A step's verification state on the canvas and in the plan panel now read from the same source, so they stay in sync after a re-plan.

Describe the workflow. MadMax builds it and tests it on your data.

Free to start, no card. Connect your apps, say what should happen in plain English, and watch each step run against real data before anything goes live.

Get started free Read the docs