Skyvern
Developer ToolsRun AI agents inside cloud browsers with Skyvern to complete tasks on websites that have no API. Agents describe a goal in plain English or run a saved agent and collect the result.
What This Integration Enables
Reading a page and acting on a page are different risks, and Skyvern is squarely on the acting side. It drives a real browser in the cloud toward a goal you describe in plain English: sign in, navigate, fill a form, submit it, download what came back. It can hold a persistent browser session so one sign in is reused across many later runs, and it can complete a login using a credential held in its own vault, Bitwarden, 1Password or Azure Key Vault without the password travelling through your flow. It can also accept a two factor code pushed in from your inbox so a waiting run can finish signing in.
Read that list again as a permissions question rather than a feature list. An agent with a session and a second factor is an agent operating with the authority of the person whose account it is using. Everything it does on that portal is attributable to them. FlowRunner's position is that this authority is exactly what makes the connector valuable and exactly what makes human-in-the-loop non optional. Skyvern gives the agent hands. FlowRunner decides which door it is allowed to open on its own. Note also that Skyvern registers no webhooks, since its webhook URL is a per run callback rather than a subscription, so a flow here starts from a schedule or an upstream system and polls Get Run for the outcome.
Without FlowRunner
With FlowRunner
Use Case Scenarios
Monthly statement retrieval from supplier portals
A finance team pulls statements from eleven supplier portals every close. The agent creates a browser session, calls Run Login Task against the vaulted credential, and once the session is signed in, calls Run Download Files Task with a navigation goal describing the documents to collect. It polls Get Run until a terminal status, reads the downloaded files list, and stores each document in S3 with the portal name and period attached. Skyvern deliberately marks a run failed when webhook delivery fails even though extraction succeeded, so the agent reads the output regardless and checks the failure reason before deciding anything is actually wrong.
Structured extraction from a portal that publishes no export
A government or banking portal shows the data on screen and offers no download. The agent calls Run Task with a plain English goal and a Data Extraction Schema, which is what turns the answer into structured JSON instead of prose. Max Steps is set on the call, because Skyvern bills per step and an agent that starts wandering should hit a ceiling rather than a budget. The extracted records go into QuickBooks Online as a reconciliation input, and the run's screenshots stay attached to the record through List Run Artifacts.
Repeatable submissions with a saved agent
Anything that runs more than once should stop being re-planned every time. The team saves the sequence as a Skyvern agent and the flow calls Run Agent instead of Run Task, with parameter keys matching the names the agent declares, read first with Get Agent. Because a saved agent's steps are pinned rather than reconsidered on each run, its behaviour is reviewable in advance. When a run fails for an environmental reason, Retry Agent Run restarts it with the parameters it originally had, which task runs have no equivalent for.
Human-in-Loop Highlight
Send TOTP Code is where this connector stops being a scraper and becomes a question of authority, so that is where the gate belongs. The run is already signed in as far as the password goes and is sitting still, waiting for the second factor that was designed specifically to require a human. FlowRunner does not treat that as a plumbing problem to be solved by piping the code out of a mailbox. The agent collects the context the person needs and asks for it: "Run rn_8841 is signing in to the Northwind supplier portal as [email protected] and its goal is to download the July statement. It is waiting on a two factor code. The code in your inbox is valid for the next few minutes. Confirm this is the run you expect, and send the code?" The controller confirms, the code goes in through Send TOTP Code, and the run continues. Two things follow from doing it this way. A code that arrives when no run is waiting becomes visible as an alarm rather than being consumed silently. And every sign in this agent performs has a named person who approved it, which is the record an auditor asks for. If the run then heads somewhere unexpected, Cancel Run and Bulk Cancel Runs are the stop, and Get Run Timeline explains what it was thinking.
Agent Capabilities
22 actionsRuns
12- Run Task Starts a one off browser task from a plain English goal and returns immediately with a run ID while Skyvern drives a cloud browser asynchronously. A Data Extraction Schema returns structured JSON rather than prose. Skyvern bills per step, so Max Steps is the main cost guard. The agent polls Get Run for the terminal status.
- Run Agent Starts a saved Skyvern agent and returns immediately with a run ID. Parameter keys must match the names declared in the agent's definition, so the agent inspects them with Get Agent first. The right choice for anything repeatable, because a saved agent's steps are pinned rather than re-planned on every run.
- Run Login Task Signs in to a website using a credential held in Skyvern's vault, Bitwarden, 1Password or Azure Key Vault, without the password passing through your flow. Paired with a browser session so the resulting signed in state is reused by later runs.
- Run Download Files Task Navigates a site and clicks through to download files, returning them on the finished run's downloaded files list. What to download is described in the navigation goal. Combined with a session created by Run Login Task when the files sit behind a sign in.
- Get Run Retrieves a task run or an agent run by ID, and is the endpoint the flow polls after starting one. The status reaches completed, failed, terminated, timed_out or canceled, at which point output holds the extracted data, downloaded files lists retrieved files, and the recording and screenshot URLs give the visual trail. A webhook delivery failure marks the run failed even when extraction succeeded, so output is read regardless.
- List Runs Returns a paginated list of runs across the account, filterable by status, run type, agent and tags. Used to sweep for failed runs, reconcile a batch of submissions, or drive a polling flow that reacts to newly completed work. Search Key needs at least three characters.
- Cancel Run Cancels a single in progress task run or agent run, moving it to canceled. Used to stop a run that is looping, heading in the wrong direction, or burning steps that were not budgeted.
- Bulk Cancel Runs Cancels up to 100 runs in one call and reports which succeeded and which could not be cancelled, typically because they had already finished. The fast route to clearing a stuck batch or shutting down runaway spend.
- Retry Agent Run Re-runs an agent run with the parameters it was originally started with, returning a new run record. Only agent runs can be retried; standalone task runs are started again with Run Task. A run not in a retryable state returns an error.
- Replay Run Webhook Resends a finished run's webhook payload. Skyvern never retries failed webhook deliveries by design, so this is the supported recovery path once the receiver is healthy again. A different URL redirects the replay somewhere else, such as a debugging endpoint.
- List Run Artifacts Returns the artifacts captured while executing a run: screenshots, screen recordings, HAR and trace files, downloaded documents and the raw model prompts and responses, each with a signed URL. Filtering by type avoids pulling the very large logging artifacts when only the final screenshot is wanted.
- Get Run Timeline Returns the nested timeline of a run, with the agent's thoughts interleaved with the blocks it executed, in order. The clearest available explanation of why a run behaved the way it did, and the first thing to read when it took an unexpected path.
Agents
3- List Agents Lists the saved agents on the account with their titles, descriptions, status and full step definitions. The agent ID field is the stable identifier passed to Run Agent. Filtering by status hides drafts, and restricting to templates browses Skyvern's prebuilt automations.
- Get Agent Retrieves a single saved agent by its permanent identifier, including the full workflow definition with every block and declared parameter. Read before calling Run Agent so the parameter keys and types are known. A version can be passed to inspect an older revision.
- Delete Agent Deletes a saved agent. Its historical runs stay queryable through List Runs, but the agent can no longer be started and disappears from listings. This cannot be undone, so the definition is exported first if it may be wanted back.
Browser Sessions
4- Create Browser Session Starts a persistent cloud browser that survives across runs, so a sign in performed once is reused by every later task. Optional ad blocker and CAPTCHA solver extensions can be loaded and the browser engine pinned. The timeout is in minutes, must be at least 5, and is capped at 240.
- List Browser Sessions Lists every browser session currently open on the account, with status, timeout and the run it is attached to. Open sessions bill continuously, so this is how a flow spots sessions left running after an earlier flow errored out.
- Get Browser Session Retrieves a single browser session, including the browser address for connecting over the Chrome DevTools Protocol and the app URL for watching the live view in the Skyvern console. Also reports files downloaded inside the session and whether live streaming is supported.
- Close Browser Session Ends a browser session and releases its cloud browser. The session cannot be reused afterwards and any signed in state is lost unless a browser profile was saved when it was created. Sessions bill for their whole lifetime, so every session a flow opens is closed.
Utilities
3- Send TOTP Code Pushes a two factor code to Skyvern so a waiting run can complete a sign in. The whole email or SMS body can be sent as the content and Skyvern extracts the code itself. The TOTP identifier must match the one given to the waiting run. Parsing the message consumes credits.
- Upload File Downloads a file from a URL you supply and uploads it into Skyvern's storage, returning a storage URI and a presigned download URL. Used to stage a document an agent needs to fill in or attach on a target site. The source URL must be publicly reachable.
- Get Server Version Returns the Skyvern server version. The endpoint needs no authentication, which makes it a cheap way to confirm the cloud API is reachable from a flow before starting an expensive run.
Frequently Asked Questions
What can FlowRunner do with Skyvern?
FlowRunner agents can run Run Task, Run Agent, and Run Login Task in Skyvern, plus 19 more actions.
Does connecting Skyvern to FlowRunner require OAuth?
No. Skyvern connects to FlowRunner with an API key, no OAuth flow required.
Can Skyvern trigger a FlowRunner workflow automatically?
Skyvern doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.
Start building with Skyvern
$100 in credits. No card required. Connect in minutes.