Skip to main content
Stay up to date with the latest changes to Cargo. This page covers the main updates shipped between February and September 2026.

​
September 2026

New features
  • cargo-ai cdk is now cargo-ai project — the command group that defines your workspace in a repo is named after what it makes. cargo-ai cdk keeps working as an alias, so nothing you have written down needs changing, and the TypeScript you import is still @cargo-ai/cdk
Integrations
  • Find email — each found address is now verified with ZeroBounce instead of Enrichley, including catch-all domains when that option is on
New features
  • Claude Code agents — Claude Code is now a runtime you pick on an agent release, alongside a repository, instead of a nested code capability. Deploy the agent from the checkout it should work on and the CDK binds the repository for you — owner, name, default branch and the project’s directory come from the git remote, with the GitHub connector taken from the project — so repository only has to be written when the agent works on some other repo. The sandbox reaches any public API rather than only Cargo, npm and GitHub, and the agent’s own shell, read and edit calls render as real cards in the chat with the command, its output and its exit code
  • cargo-ai ask — a chat in your terminal, running Claude Code against the working copy on disk, so it reads and edits the files you already have. ask "…" opens with a first message, --continue resumes the last chat, --print runs a single turn as JSON, and typing prose into the command palette sends it here. Anything that touches the machine asks for approval first, the spinner names the tool that is running, and a stalled turn recovers into a new session instead of dropping you back at the shell. CLI chats are their own kind of activity in the app
  • Email-driven agents — an agent can now be woken by its own mail: the first sendEmail from a chat binds it to that thread, and a later open, click, reply, bounce or unsubscribe starts the next turn in every chat on the thread. The agent reads the history with a new listEmailEvents action and ends the conversation for good with a close-chat tool, so a follow-up sequence stops when it should. Native triggers are typed in the CDK through agentNativeTrigger, and an event no longer has to wait out a pending heartbeat before it is answered
  • Connect a domain you already own — Settings > Domains gains Connect domain next to Buy domain: publish the TXT record it shows you and the domain verifies itself, at no cost, instead of only being able to send from a name bought through Cargo. Mailboxes on a connected domain are not available yet
  • Capabilities on MCP servers — a workspace MCP server can expose built-in abilities such as web search and workspace context as tools of its own, set from the server editor, the API, the CLI or defineMcpServer({ capabilities: [...] }). Web search no longer needs a chat behind it, so it works over MCP
  • Home as one setup path — the app’s Home is a single numbered sequence — install the CLI, sign in, create the project, add a cookbook, read the plan, deploy — in a terminal flavour and a coding-agent flavour, and picking a cookbook rewrites every command below it. The workspace panel counts all six deployable resources and links to Traces and Alerts
  • cdk init sets up git and GitHub — after the project is written, init runs git init, offers to create a private GitHub repository, and offers to connect GitHub to Cargo and write the connector file, so the next steps only list what is actually left to do
  • cargo-ai book-demo — open the same booking form as Talk to sales from the terminal, with your account email prefilled; --no-open prints the URL for scripts and agents
Integrations
  • X — getFollowers and getFollowing read up to 75,000 accounts in a single call, up from 10,000. Every people-returning action returns handle, name, location, avatar and account creation date again, after X moved those fields in its payload
  • ChatGPT — connecting Cargo’s MCP server from ChatGPT or Codex completes, since the per-connector OAuth callback they use is now accepted. Every platform MCP tool advertises an output schema and returns structured content, and run, batch and list responses carry the result rather than internal debug payloads
  • Anthropic — an instruct call with web search returns the whole answer instead of the fragment left after the last citation
  • SendGrid — sendEmail fills the variables of a dynamic template again, rather than sending the template blank
  • Detect job change — the action declares its output, so job_change_status and the person’s fields can be picked in a downstream node
Improvements
  • query_models over MCP takes a read-only SELECT across dataset.model references, so a client can count, filter and join instead of paging a whole model through the tool. modelUuid and offset are gone, and query is required
  • An MCP session left idle between calls stays open instead of failing the next one with a closed connection
  • An agent that queries a model with timestamp or date columns carries on to the next step instead of failing on the result
  • Asking an agent to analyse a file no longer puts the chat into an error state
  • Hosted app builds run to the full 18-minute budget instead of being killed at five minutes and reported as an internal error with no build log
  • cargo-cdk plan, check and deploy work from any subdirectory of the project, and deploy stops before applying when a resource still needs a secret() that is missing from your .env
  • Deploying a defineMailbox no longer waits for the mailbox to become active, so a graph deploys while the mailbox is still pending
  • HTML written by an agent arrives as formatted mail instead of visible tags
  • An agent draft saves when no model provider is set, which is the normal state for a Claude Code agent
  • A play on a watched-records trigger starts its dependent plays once the batch finishes
  • An action config pinned to a number on a release is enforced on the agent’s tool schema, so the model cannot ask for a different value
  • Chats and runs with nothing in flight are labelled Idle rather than spinning, and the Generating filter matches only what is really generating
  • The Activity field on Wait for email event is a set of radio options again instead of an expression input
  • Native actions no longer show as unknown icons in the agents list, and the Tools and Integrations columns are both called Actions
  • Running a node through the API requires workflow write permission, so a viewer can no longer execute one
  • Installing Cargo from a coding agent is four commands from a guide served at api.getcargo.io/INSTALL.md; the install.sh bootstrap that wrote into your Claude Code config is retired
  • Adding a cookbook ends on a setup checklist naming what to look up in the workspace and what only you can answer, and its scripts are installed under scripts/<name>/

​
August 2026

New features
  • Platform MCP server — connect ChatGPT, Claude or Cursor to mcp.getcargo.io/mcp and operate the whole workspace without publishing a server of your own: find, price and execute any action, run tools, ask agents, query a model’s records, list plays, runs and connectors, and check credit usage. If you have a Cargo connector, enrichment and search tools come with it. The server sends a client its operating instructions on connect, so an assistant knows how the tools fit together with nothing to install, and cargo-ai mcp bridges it with no --server
  • Workspace context over MCP — a connected client reads the playbooks, ICP definitions and rubrics your workspace keeps as context, through map, browse, search and read calls, so it answers from what your team actually wrote instead of what the model assumes. Read only, with no write, edit or shell access
  • A CLI that opens on what you want to do — running cargo-ai at a terminal signs you in, lands you in a workspace and offers a destination: open a coding agent, run the guided demo, scaffold a project, or explore. Returning users get a searchable command palette where an option that wants an id offers the workspace’s own rows, an option with a fixed set offers a menu, and a table can be narrowed in place with the filters the command already declares. Setup problems print the command that fixes them. The whole flow also runs headless with flags, and anything non-interactive behaves exactly as before
  • Connectors from the CLI — cargo-ai connection connector create and update open the real connection form in a browser, handle the OAuth consent there, then hand you back to the terminal on a done screen, and cargo-ai cdk add connector <slug> also writes the connector file and binds it to your project state. Defining an OAuth integration in code no longer needs a manual detour through the app
  • cargo-cdk info — the resolved project on one screen and offline: every resource in dependency order, the file it was declared in, whether it is new, live, draft or orphaned, what it depends on, and any problem blocking a deploy. --json emits the whole report
  • Cookbooks in an existing project — cargo-cdk catalog list, search and view browse the cookbook catalog and cargo-cdk add cookbook/<name> installs one into a project you already have, instead of only into an empty directory. cdk init is interactive, takes --cookbook, and seeds your company context from your domain
  • Wait for email event — a new native action holds a play until an open, click, reply or unsubscribe lands anywhere on the thread of an email you sent, or until the timeout you set
  • Mailboxes on members — attach mailboxes to a revenue-organization member, from member settings, from defineMember in the CDK, or with --mailbox-uuids on the CLI. Each mailbox belongs to one member
Integrations
  • OpenAI — added the GPT-5.6 family, Sol, Terra and Luna, at the top of the model picker
  • Surfe — new find person by email action, returning name, company, job title, country and LinkedIn URL
  • LinkedIn — profile enrichment returns country as a name rather than a code, a profile with no location enriches instead of failing, and a company join date is derived from the whole run at that company. Post search accepts company IDs again, and the extension reconnects its identity on every sync
  • AI Ark — gateway timeouts are retried for up to three attempts instead of failing the run, and an exhausted retry reports a short error rather than the provider’s raw response
  • Google Sheets — accented column headers are normalized, so a sheet that uses them syncs instead of failing
  • HubSpot — sequence enrollment accepts numeric IDs
  • Lusha — a node fails with the reason when the Lusha account is out of credits, instead of reporting an empty result
  • Explorium — saving listen events skips ids the provider does not match, rather than failing the save
Improvements
  • Plays and tools deployed from the CLI or the CDK are laid out automatically, so a graph written in code opens readable in the editor
  • orchestration action list returns a typed credit price for every billed action, natives included, showing whether it is fixed, metered per unit or sold in packages, and documenting that a fixed cost adds to the metered rate rather than replacing it
  • orchestration action get-output-schema accepts the same action object action list hands back, plus --data for the inputs that decide the shape, so the examples in its own help work as written
  • An email open is no longer recorded for a security scanner or for a pixel fetched within ten seconds of sending, so a wait-on-open never fires on a prefetch
  • Filter values on Match record and Search records nodes open in expression mode, so / shows the previous node’s values
  • Renaming a node in the editor keeps the custom name when you click its label
  • Exporting run outputs completes when one stored object is not valid JSON, instead of failing the whole export
  • Connecting an MCP server that authenticates with client_secret_basic works
  • API token secrets are masked to the last four characters with reveal and copy, and the action docs, MCP server and ingest pages point at an existing token instead of minting a “Quick access” one
  • The Mailboxes list stops showing every mailbox as updated a minute ago while replies sync
  • Create and Save sit in a full-width footer on the connector, model, tool, play, worker, app and agent edit screens
New features
  • Mailboxes — provision a fleet of sending mailboxes on a domain you own, warm them up toward a 40-a-day target, and send from a play, a tool or an agent with the new sendEmail action, which paces itself to whatever that mailbox’s allowance is on the day. Every message carries one-click unsubscribe, opt-outs land on a suppression list checked on every send, and mailboxes are billed monthly
  • Custom roles — Settings > Roles lets an admin build a role from the same per-domain Disabled / Read / Read & write matrix as API tokens, then assign it to members. The four built-in roles are unchanged
  • Remote MCP clients over OAuth — ChatGPT, Claude and Cursor authorize against Cargo itself, so connecting one is a consent screen instead of a pasted token, and an admin can see every connected app on the MCP server page and revoke it. Agents find the server from a card published at /.well-known/mcp
  • Analytics and logs for hosted apps and workers — an Analytics tab on an app lists ingested events and a Logs tab on a worker lists its runtime output, both filterable by search, kind or level and date range. Every worker invocation is also recorded as a trace, with the worker’s console output under its span
  • Agent-ready REST API — every 4xx and 5xx returns a typed error body, every operation carries a stable id, rate-limit headers ride on every /v1 response including 401s, writes accept an Idempotency-Key so a retried request replays rather than duplicating, and long-running creates return 202 Accepted with a Location to poll. The versioning and deprecation policy is published alongside the reference
  • Synchronous action execution — waitUntilFinished on action execute and execute-batch holds the request until the run or batch reaches a terminal status, up to five minutes, and returns the finished run with its context. Anything slower comes back as a 408 carrying the run, so you can keep polling
Integrations
  • HubSpot — run a Cargo tool from inside HubSpot, either through the new Run Cargo tool workflow action, which enrols a record and reports back once the run finishes, or from a Cargo tools card pinned to a contact or company. Reconnect your HubSpot connector to pick up the scope the workflow callback needs
  • Attio — new merge action, combining two records into one through Attio’s record merge endpoint
  • LinkedIn — post extraction takes a limit of 20 to 100 posts and returns them in a single pass, which restores extraction after the provider dropped the pagination it relied on. Backfilling further back than that limit is no longer possible
  • Sillage — signal detection advances past each batch instead of re-fetching and re-billing the same signals every hour, a persona saved without a location is caught on the form with a readable message instead of failing the model, and copying a listen-signals model gives the copy its own workspace
Improvements
  • Runs, records, spans and chats have a search box matching on title, id and, for spans, the execution error, with --search on the matching CLI commands and no % wildcards to write
  • --is-enabled false and every other boolean CLI flag mean what they say, instead of being read as the flag on its own
  • Deleting records no longer requires available credits, so a workspace over quota can still clean up
  • Model records larger than 3 MiB are rejected with a clear error when written, instead of failing the batch further down
  • Large warehouse-backed model and segment pages load without exhausting query memory
  • A script node that finished successfully no longer reports itself as failed
  • Clicking anywhere on a row in the Plays, Agents, Apps, Tools, Workers, Models, Alerts, MCP servers and Files lists opens it
  • Observe sits above Connect in the workspace sidebar
  • Tool cards no longer show a credit range of “max. NaN”
  • Workspace creation is capped at three per user per day
  • cargo-ai manifest add is removed: cookbook folders now ship as skills a coding agent places in your project, and cargo-ai cdk init --from stays as a plain scaffold
New features
  • JavaScript models — defineScript fills a model from code you write in the CDK: the script runs in a sandbox on every sync, reads its own encrypted secrets from the model’s env, and returns the rows that become the table, so anything you can reach in code becomes a model
  • Node graph diagrams — cargo-ai orchestration node diagram draws a play, tool, release or run as a Mermaid flowchart or as ASCII for a terminal, marking the steps that bill, the fallback paths, and any node unreachable from start. It runs nothing and costs no credits
  • Models and files as agent capabilities — attach as many models and files to an agent as you like without growing its tool list, and read a file exactly through new list, search and read calls instead of only searching it as a vector. Bulk-allowed actions are now one tool taking an items array, and a single item comes back as a record rather than a one-row CSV
  • Custom column writes — the Model insert, update and upsert nodes map custom columns alongside original ones, so writing one no longer needs a node of its own, and POST /storage/columns/insertCustomValues and cargo-ai storage column insert-custom-values do the same from the API and the CLI
  • Unattended payment methods — an agent can put a card on file with nobody watching, through a link that opens on the card form itself or by passing a payment method tokenized elsewhere. A card that cannot be charged is reported when you add it, with the issuer’s decline reason, instead of at the next renewal
  • Action reference in the app — Connect > Actions lists native actions such as modelSearch, modelAsk, scoring and python alongside agents, tools and integration actions, and documents every one of them through the same execute endpoints, with a cURL, @cargo-ai/api and cargo-ai switcher on each snippet
  • Browserless setup for coding agents — npx skills add getcargo.ai or docs.getcargo.ai installs the Cargo skill bundle, install.sh takes CARGO_EMAIL and CARGO_CODE to create the account and sign in with no browser at any point, and both domains serve /llms-full.txt
Integrations
  • Datagma — new enrich person action keyed on a LinkedIn URL or a professional email, plus a second one that resolves a profile from a personal email address
  • Apollo — runs first in the Find email and Find LinkedIn URL from email waterfalls, and a new Find LinkedIn company URL from domain tool ships the full provider waterfall
  • Reverse Contact — moved to the provider’s V2 API, which restores every action. Reconnect with a V2 key; keys issued for V1 no longer authenticate
  • Wiza — enrichment waits for the reveal to finish, so a profile it does find is no longer reported as a miss, and a connector left without an API key says so instead of reporting itself connected
  • Sales Navigator — a run that names one of your own LinkedIn users spends that seat’s allowance instead of being metered against the shared pool quota
  • HubSpot — a private-app token is verified when you save the connection, and the merge action accepts numeric primary IDs
  • Custom integrations — a worker outage no longer wipes a connector’s registered actions, since the last good manifest is served until the worker recovers. Valid extractor responses are accepted, and a webhook record that leaves a field empty still ingests
  • FullEnrich — reverse email lookup completes on a webhook
  • CompanyEnrich — nothing is billed when the lookup finds nothing
  • Metabase — the column autocomplete accepts a numeric question id
Improvements
  • Runs start faster when many runs are already open
  • cargo-ai orchestration run download takes --uuids to export a specific set of runs, and --is-finished is accepted again on run list, count and download
  • A failed model sync or run reports the underlying error instead of a generic message
  • Filter builders label the second button “Add group” once a group exists, and a saved segment hides Reset | Update after you update it
  • The run, node and batch inspectors stay open through play, tool and agent redirects
  • Nested groups are rejected on import, validate and deploy, with a clear error
  • Signing in works where the home directory is read-only, honouring CARGO_CONFIG_DIR and XDG_CONFIG_HOME
New features
  • Traces — a new Traces section under Observe: filter traces by trigger, status and date, then open one as a flame graph, a waterfall or a span list, with the inputs, outputs and config of the selected node alongside it
  • Alerts — schedule an alert on spans, on whole runs, on read-only SQL against your warehouse, or on a model’s own health (record count, share of a filter, freshness since the last sync, last sync duration), and have it notify through Slack, Teams, Resend, SendGrid or an HTTP call
  • Batch health alerts run any action — a play’s batch health alert is no longer limited to one Slack message: it can fire any action, you can configure several, and each one now executes as a real run with its own trace, retries and credit accounting
  • Sign in without a per-machine token — cargo-ai login --email signs you in with a code sent to your address, so it works from a CI container or an agent shell with no browser. The session covers every workspace you belong to instead of being pinned to one, cargo-cdk renews it from the same credential, and cargo-ai logout revokes it for good
  • cargo-ai billing add-payment-method — attach a card from the CLI, the last workspace setup step that still required the dashboard
  • defineMember — reference a revenue-org member by email in a CDK spec, so a repo that routes leads through territories is no longer pinned to one workspace’s member ids
  • defineDomain — register a sending domain through Cargo from a CDK spec and publish its DNS zone alongside it
Integrations
  • Metabase — new integration: run a saved question, run SQL against a connected database, search entities, sync a question’s rows into a model, and hand an agent Metabase’s own MCP tools. An admin must enable the MCP server under Admin > AI > MCP first
  • Replicas — new integration with a launch agent action that runs a coding agent in a cloud workspace and returns its branches, diffs and pull requests to the next node
  • Devin — new integration with a launch agent action that returns Devin’s structured output and pull requests
  • AI Ark — new enrich company action, keyed on a domain or LinkedIn URL at 0.01 credits, plus country and financial fields on company records and a search limit that shows the default you get
  • LinkedIn and Sales Navigator — teammates who synced their session through the Chrome extension now show up as Revenue Organization members, so a capacity can cap their connection requests, messages and profile visits the way it caps CRM records. The Linkedin Users fields now link to the extension
  • GitHub — connections request the scopes needed to reach private organization repositories, so they appear in the repository picker and the code capability can clone them and open pull requests. Reconnect an existing GitHub connector to pick this up
Improvements
  • Sales Navigator usage is metered per workspace against quota buckets, 1,000 requests per bucket per rolling 24 hours, so one workspace’s volume no longer eats the allowance everyone shares
  • Exporting run and record outputs streams to storage, so large exports complete instead of failing
  • Warehouse queries are generated without redundant work, cutting the time segment filters, record lookups and record counts spend on your warehouse
  • Opening a play or a tool no longer freezes the tab
  • Intermittent “failed to fetch” errors on long-lived browser sessions are fixed
  • The run, node, batch, chat, message and file inspectors open in a side drawer, so they stack cleanly instead of dropping a dialog over the trace panel
  • Writing to a column a model does not have is rejected naming the unknown and the available columns, instead of succeeding and silently dropping the value
  • Segment create and update return the synced record count and sync time, instead of always reporting zero records
  • JavaScript script nodes accept a result variable as their output, matching Python nodes, and the editor ships with a starter example
  • Creating a batch from a play’s generated segment is refused with a message that says so, instead of reporting no records
  • Agent descriptions longer than 255 characters save
  • A fixed-price action bills nothing when it finds nothing, so a miss on AI Ark enrichment is no longer charged as a hit
  • A model refresh that is already running reports itself as refreshing instead of failing with a lock error
  • Span search matches on the execution error message as well as the title
  • API tokens can be granted Observability, so a token can read workspace events and manage alerts
  • The workspace menu links to the Cargo status page
  • Phantombuster models backfill an agent’s full run history instead of only its most recent containers
  • FullEnrich requests are paced instead of released in bursts, which stops the 429s and the enrichment requests that could never be satisfied
  • Sillage connections surface configuration problems as readable rejections, sync when a top account has no removable identifier, and stop their agents when a model resolves no account
  • LinkedIn matched audiences are created with a valid source platform
  • The webhook token alert is hidden on native models
  • Workspaces with very large numbers of models sync faster

​
July 2026

New features
  • cdk deploy --draft — stage plays, tools, agents, workers, and apps as unpublished drafts, test them, then publish with a plain cdk deploy
  • Managed play on/off switch — definePlay takes isEnabled, so a play ships enabled or held off from code instead of needing a manual toggle
  • Model config redeploys — a deploy that changes a model’s config relaunches a reset run, so the data is rebuilt under the new config
  • Action search and doctor — cargo-ai connection action search finds any integration action by keyword, and cargo-ai doctor reports CLI version, credentials, and API reachability in one command
  • Project .env auto-discovery — the CLI and CDK pick up CARGO_API_TOKEN, CARGO_WORKSPACE_UUID, and CARGO_BASE_URL from the nearest project .env, so a repo pinned to one workspace stays on it
  • Snowflake statement timeout — set statementTimeoutSeconds on a Snowflake source to cancel runaway queries server-side
  • Browser tab titles — tabs show the current page or entity name instead of “Cargo - App”, and keep it when you filter or sort
Integrations
  • LinkedIn — new Enrich Post and Message Profile actions, new page followers, followers, and company viewers extractors, and post comment and reaction search now accepts a post URL
  • AI Ark — new count people and count companies actions, searches paginate up to 2,000 records, and the Technologies and NAICS filters have searchable dropdowns
  • Gemini — added the Gemini 3.6 Flash and 3.5 Flash-Lite models, with 3.6 Flash as the new default
  • Apollo.io — clearer sequence options and contact mapping fields
Improvements
  • Haiku is now priced at the small-model tier, 0.05 credits per 1K tokens
  • Agents default to a temperature of 1, so releases on current Anthropic models no longer error
  • Model schedules are validated against the extractor’s minimum interval, instead of scheduling runs that can never do anything
  • Deleting a field from a CDK spec now clears it on the live resource, and a deploy creates resources or fails rather than silently binding to one it did not create
  • Model tables keep refreshing while a workspace is out of credits
  • Model search no longer waits behind a table refresh, so high-volume searches stop failing on lock timeouts
  • Sorting records on a column the model’s table does not have is rejected with a clear message
  • You can sign in from a headless environment, such as a CI container or an agent session, through the device flow
  • Tool nodes report their real result when the final workflow node is not named end
  • A Sillage connection with no top accounts syncs instead of failing
New features
  • Code capability — from an agent chat, delegate coding work on your own GitHub repository to a coding agent running in a sandbox, and get a pull request back
  • Generate an output schema from a prompt — write the prompt, hit Generate, and the structured-output JSON schema is filled in for you
  • Large file uploads — files upload straight to S3 in parallel parts, so multi-gigabyte CSVs no longer break mid-transfer
  • Shareable model views — the selected segment, filters, sort, record limit, and visible columns live in the URL, so a refresh or a shared link restores the same view
  • API token descriptions — tokens take an optional description, and the auto-created token is now “Quick access”
  • Event-only entities — a company or person seen only in events now gets a row in unified Accounts and Contacts, so its joined columns resolve
Integrations
  • AI Ark — new integration with six actions and two extractors, covering search, profile enrichment, email and mobile lookup, and reverse lookup
  • X (Twitter) — new connector with user, post, and search actions plus matching extractors
  • BuiltWith — new enrichment integration with enrich domain and find websites by technology
  • OpenRouter — new language-model provider, so agents can run any model OpenRouter fronts
  • Parallel and Exa — billable with Cargo credits, so they run without a key of your own
  • Apollo — expanded search, new people and organization extractors, and location autocomplete on company search and fetch
  • Sillage — new fetch leads extractor and search leads action, signal agents in the extractor, and source columns on signals
  • Dropcontact — find email now runs through async webhooks
  • HubSpot — isEmpty and isNotEmpty filters, and properties whose internal name contains ___ are kept
  • Surfe — fixed authentication and enrichment webhook handling
  • Smartlead — fixed listen-events webhook registration and webhook deletion
  • OpenAI — removed the invalid gpt-5.5-mini and gpt-5.5-nano models
Improvements
  • cdk plan loads resources about 50x faster on a large workspace
  • Worker environment variables resolve at runtime, and secrets stay out of the deployed bundle
  • The enroll modal starts with no records selected
  • Each unification strength option explains what it does
  • Segment fetch and stream return records in a deterministic order
  • Run errors name the reason a run could not be created
  • Empty filter groups are rejected instead of matching everything
  • A tool release with no form fields is refused at deploy time
  • Number columns cast to dates infer their epoch unit
New features
  • CDK manifests — new cdk manifest init and cdk manifest add commands to set up GTM repos
  • Test before you deploy — run agents and workflows straight from your CDK project, plus cdk check, slug hints, and loop DX
  • cargo-ai mcp — a stdio bridge that connects any MCP client to your Cargo MCP servers
  • Local worker development — npm run dev in the worker SDK for a local dev loop
  • Realtime computed columns — computed columns are evaluated on the fly for realtime plays
  • Filtered watch schedules — watch schedules can now run on models with a config filter
  • Download deployment source — grab the source of any hosting deployment from its dropdown
  • Default API token — workspaces now surface a user-facing default token
  • Per-reference unification — configure unification per reference, with account slug matching
Integrations
  • Anthropic — added the Claude Sonnet 5 model
  • Salesforce — REST request and invoke flow actions
  • Lemlist — CRM contact and company actions
  • Company Enrich — Get Workforce and Lookup Person actions
  • Slack — clearer agent message display
Improvements
  • Token usage is now returned by OpenAI and Anthropic instruct actions
  • Chat renders every context capability tool invocation
  • Segment filters are validated against column types
  • The end node is now a regular node you can add and remove
  • Territory descriptions accept long-form text
New features
  • cdk pull — generate define* source from a live workspace
  • cdk init --from — scaffold a project from a cookbook repo on GitHub
  • Models in the CDK — native object and unified models, unification, relationships, and typed model.columns accessors
  • Typed action outputs — integration action outputs are typed in workflow codegen
  • App capability — agents can edit a hosted app in a git-backed sandbox runtime
  • Worker cron triggers — schedule workers, with custom-integration auto-detect and automatic Cargo API auth
  • Export deployment source — download any deployment as a tar.gz
  • cargo-quickstart — a guided first-run demo launched by the installer
Integrations
  • Sillage — new integration with a listen signals extractor
  • Default connector per integration — pick which connector an integration uses by default
Improvements
  • MCP tool calls no longer time out on long-running actions
  • Computed column DDL is idempotent and surfaces structured errors
  • Record creation returns a readable message on validation errors
  • Manual runs in the play editor no longer hang or spawn duplicate runs
New features
  • Cargo CDK — define your entire Cargo setup in code and deploy it, including capacities and territories
  • Standalone cargo-cdk bin — run the CDK as its own command
  • Single sign-on per workspace — enforce a login provider per workspace, with Okta SSO support
  • Edit in editor — reopen a past run in the editor to reproduce and iterate on it
  • Inline file viewer — preview files directly in the files library
Integrations
  • Surfe — new integration
  • LinkedIn & Sales Navigator — expired identities are now flagged in connection pickers so you can reconnect
Improvements
  • HTTP webhook triggers now support insert, update, and remove events
  • Renamed “Revenue Organization” to “Territories” across the product
  • Restrict workspace access to company email domains
  • Workers are now built from TypeScript source server-side

​
June 2026

New features
  • Parallel Create Task — new Parallel AI action with webhook support
  • Smarter exports — exports now respect your row selection and use meaningful filenames
Integrations
  • Perplexity — now selectable as an agent model
  • Salesforce — hosted MCP integration
  • LinkedIn — new Extract Similar Companies action
Improvements
  • Raised the HTTP Call action timeout to 6 minutes for slower endpoints
  • Faster, more reliable runs across the platform
New features
  • Agent output evaluator — automatically grade and gate agent output quality
  • Resumable batches — large runs batches now resume where they left off
  • Workflow deploy diff — review changes before deploying a workflow
  • Per-segment columns — your column selection now persists per segment
Integrations
  • HubSpot & Notion — dedicated MCP integrations
  • LinkedIn — Extract Company Employees Insights action
  • Sales Navigator — person and company search metrics
  • Lemlist — inbox and task actions
  • Bright Data — Facebook profile scraping and async social actions
Improvements
  • Content permissions can now be granted per role
New features
  • Workflow SDK — define workflows in code with TypeScript
  • Domain management — buy and manage sending domains in-app, powered by the new Mailpool integration
  • Cargo Forms — embeddable forms with a public form trigger
  • Hosting previews — per-deployment preview URLs for apps and workers, plus custom domains
  • Run date filters — filter runs by updated-after / updated-before
Integrations
  • Bright Data — Instagram, TikTok, Facebook, X, and YouTube scraping actions
  • Salesloft — create notes on persons and accounts
  • Anthropic — added the Claude Fable 5 model
Improvements
  • Context files auto-fill missing title and description
New features
  • Agents write to data models — agents can insert, update, upsert, and delete records
  • Content library — files organized into a dedicated Content area with libraries
  • Model relationships — link native object models together
  • Expandable record view — richer record modal with model icons and a Canvas view toggle
  • CLI update notifier — the CLI now tells you when a new version is available
Integrations
  • Parallel AI — new integration with Search and Extract actions
  • Google Ads — Customer Match audiences
  • Redshift — support for views in the fetch table extractor
Improvements
  • Row-level filters on custom unification
  • Link-shortener domains are now blocked during unification for cleaner data

​
May 2026

New features
  • Agent memory — agents can now retain memory across conversations
  • Chat heartbeat — conversations auto-wake to continue long-running work
  • Code interpreter & file attachments — improved code interpreter and file attachments in chat
  • File search — search across your uploaded files
  • Preview action outputs — see an action’s output schema without running it
  • Wait for completion — new option to wait for tool/agent steps to finish
Integrations
  • OpenAI & Anthropic — added GPT 5.5 / 5.3 and Claude Opus 4.8 models
New features
  • Structured AI output — define a JSON schema for AI output at the release level
  • Unification column selection — choose which columns to unify
  • Human review in Slack — review items in Slack with an optional Edit button
Integrations
  • Airscale — personal email, reverse email, reverse phone, and Find People
  • Smartlead — webhook events extractor
  • HeyReach — list users for revenue organization sync
Improvements
  • Integration models can now live inside model folders
  • Choose a default output type when creating AI messages
New features
  • LinkedIn profile viewers — incremental viewers with a configurable extract limit
Integrations
  • Airscale — new integration with a phone lookup action
  • Explorium — expanded business search filters
  • NetSuite — improved field schema handling
New features
  • Apps & Workers hosting — host apps and workers directly on Cargo
  • Context files — new Context files and Context Agent
  • Edit records inline — create and edit model records directly from the model view
  • Monthly schedules — added an ‘Every month’ scheduling option
  • Longer agent runs — raised the AI agent max steps from 30 to 100
Integrations
  • Marketo — new integration with Custom Objects support
  • Mixpanel — EU data residency support
  • Outreach — richer event attributes
  • Cargo — prospect and business event triggers
Improvements
  • Raised waterfall throughput for faster enrichment

​
April 2026

New features
  • LinkedIn Matched Audiences — sync matched audiences via Polytomic
  • LinkedIn profile viewers — fetch and extract profile viewers
  • SQL exports — download query results directly
  • Persistent filters — Logs and Agent Activities filters now persist in the URL
Improvements
  • Renamed orchestration logs to “spans” for clarity
New features
  • Scoped API tokens — create tokens with named, granular permissions
  • Query your runtime data — run SQL over runs, batches, and records
  • Easier CLI setup — CLI device-flow login and a one-line installer
  • Workspace reports — submit feedback reports from your workspace
Integrations
  • Anthropic — added Claude Opus 7
  • Cargo — Bombora intent topics enrichment
  • LinkedIn — profile activity extraction
Improvements
  • Promoted Claude Code as the primary install path
  • Raised the sync cap to 12k records for BigQuery and Snowflake
New features
  • Instant record updates — record changes now apply immediately
  • Concurrent runs limit — set a per-workspace concurrent runs limit
  • Node status filter — filter model executions by node execution status
Integrations
  • OpenAI — added GPT 5.4 mini and GPT 5.4 nano models
Improvements
  • Billing settings actions now respect write permissions
New features
  • Tools are now “actions” — consistent terminology across the app
  • Orchestration root nodes — with last-execution outputs
  • Self-serve tier — new 1.5K credits/mo plan
New features
  • MCP resources & templates — resources support and reusable server templates for MCP
  • Redesigned model list — grouped objects with active/inactive datasets
  • Workspace logs view — see logs and execution timestamps across your workspace
  • Signed batch webhooks — secret signing for batch completion webhooks
Integrations
  • Sales Navigator — company search metrics
Improvements
  • Added unification mappings for official integrations
  • Renamed “Storage” to “Data Model”
  • Clearer CLI help text and a revamped sidebar

​
March 2026

New features
  • Record outputs export — export the output of individual records from workflow runs
  • Date filters on runs — filter workflow run history by date range
  • Monitoring webhooks — receive webhook notifications for workflow monitoring events
Integrations
  • Piloterr — added G2 product info endpoint integration
  • Waterfall — added logging for waterfall webhooks
Improvements
  • AI chats now use a table layout for better readability
  • Schedule display fixes
New features
  • Community templates — browse and use community-contributed workflow templates
  • Schedule on ingest model — schedule-based triggers now supported on ingest models
  • Open tool/agent button — quickly jump to a tool or agent definition from the AI tools editor
Integrations
  • Email Bison — new integration with upsert lead action
  • PhantomBuster — added fetch agent results action and UI schema
  • LinkedIn — fixed post comment and reaction extraction
New features
  • GPT-5.4 model support — added GPT-5.4 as an available language model
  • Sales Navigator fetch accounts — new action to fetch account data from Sales Navigator with account-level filters
  • Context resources — introduce contextual data resources for workflows
  • Autocomplete search — search-on-change support for autocomplete fields across integrations
Integrations
  • Instantly — added search action
  • Sales Navigator — account filters and extractor UUID support
Improvements
  • Pricing page now shows enrichment credit details
  • Faster BigQuery and Snowflake downloads
  • Improved node validation
  • Improved model pause behavior
New features
  • Integration MCP clients — integrations can now expose MCP-compatible tool clients
  • Sales Navigator search limit — raised search leads limit to 25,000 records
  • Release version validation — added validation for release versions
Integrations
  • Resend — new Resend API integration for transactional email
  • Adyntel — new Adyntel integration
  • DataFast — new DataFast integration with auto page view capture
  • PhantomBuster — added fetch agent results action
  • Firecrawl — added search endpoint
  • Icypeas — added findPeople and findCompanies actions
  • EnrichCRM — added getFunding action
Improvements
  • Improved Sales Navigator search
  • Improved “let agent decide” mode
  • URL verification webhook support

​
February 2026

New features
  • Burst rate limiting — introduced burst rate limit support for better throughput control
  • Webhooks for waterfall and full enrich — webhook-based event delivery for waterfall and full enrich workflows
  • Multi-dataset model relationships — models now support relationships across multiple datasets
  • Package publishing through app — publish npm packages directly through the Cargo app
Integrations
  • Attio — added note creation and list notes actions
  • Cognism — improved enrichment with chunked processing
  • Snowflake — added batching support for actions
  • BigQuery — added batch support
Improvements
  • Increased LinkedIn rate limits
  • Added sync after model creation
  • Workflow termination on model deletion
  • Updated email waterfall and Salesforce dedup templates
New features
  • Cargo SDK — introduced the official Cargo SDK for programmatic access
  • Cargo CLI — introduced the Cargo command-line interface
  • Waiting mode — implemented waiting mode for long-running workflow steps
  • Message cancel — support for cancelling in-progress AI messages
  • LinkedIn event attendees — new capability to extract LinkedIn event attendees
Integrations
  • Exa — new Exa search integration
  • Gemini — added Gemini models with credit support
  • Anthropic — added new Anthropic model variants
Improvements
  • LinkedIn find profile URL capability
  • Human review now tracks user assignment and logs
  • File upload body limit increased to 2GB
  • Cancel downstream runs when parent is cancelled
  • Improved batch download performance
  • Renamed “agent deployment” to “trigger” across the UI
New features
  • AI rate limiting — implemented rate limiting for AI language models and AI worker activities
  • Node modal — new modal interface for node configuration
  • Model DDL — introduced data definition language support for models
  • SoR readonly queries — system of record now supports read-only query mode
  • Enrollment filter — new filter type for enrollment-based targeting
  • Trigger rename — renamed “tool launch” and “agent deployment” to “trigger” for consistency
Improvements
  • More reliable model sync
  • LinkedIn find profile URL action
  • Improved segment filter validation
  • Credits support for company enrichment
  • Improved LinkedIn company enrichment coverage from domain
New features
  • Lookup columns — introduced lookup columns for cross-model data references
  • Recipes — re-introduced the recipes system for reusable workflow patterns
  • Unification property modal — new modal for managing unification property mappings
  • Confirmation modal for scheduled triggers — added a confirmation step when switching to scheduled trigger mode
Integrations
  • Anthropic — added Opus 4.6 model support
  • OpenAI — added missing model variants
  • Outreach — updated integration with opportunity prospect role and event types
  • Apollo — updated search endpoint
Improvements
  • Faster batch starts
  • Faster agent execution
  • Stripe subscription cancellation support
  • Upload limit increased to 500MB
New features
  • Integration language models — integrations can now expose their own language models for use in AI actions
  • Bearer token auth — API now supports bearer token authentication
Integrations
  • Outreach — added event name selection for Outreach events
Improvements
  • Improved file upload handling
  • Added description tooltips on file widget