Activepieces

By the whole run, not by the step. The pricing FAQ says a run is one credit however many steps it has, and AI is the exception on top: an AI step costs 2 credits on a fast model, 10 on a smart model and 20 on a frontier model, and each agentic action is 1 credit. Connect your own AI provider key on Plus or above and every AI step drops to 1 credit, with the model cost moving to your provider bill.

What happens at the limit depends on the plan. On Free, flows stay switched on but runs past the quota do not go through until the credits refresh. On Plus and Team they keep running and the extra is billed at $0.007 per credit. The page is not consistent about Free: its comparison table lists a daily reset and “waits for tomorrow”, while the FAQ describes 1,000 credits a month and waiting until next month. Check the counter in your account before you rely on a free daily allowance.

Hosting, not features. The licence page puts the core under MIT and the enterprise directories under a commercial licence. Community Edition is that core on its own: the pricing FAQ says paying adds access control, change history, separate projects, safe releases and running Agents and Chat. The license key docs add that every plan on the pricing page is available self-hosted, and Plus or Team subscribed in the app applies without a key.

The recommended install is one command that starts four containers (app, worker, PostgreSQL, Redis) and needs Docker Compose v2, at least 2 vCPU and 4 GB of RAM. Back up the generated .env file: without its encryption key, stored connections cannot be decrypted even from a full database backup. The single-container hobbyist install runs on an embedded database and cannot move to Enterprise later. If you are weighing self-hosted tools by billing unit, the n8n profile sets out its Community edition and per-execution cloud plans.

A flow is a trigger followed by the steps you choose, and it runs the same way every time. An agent is briefed in plain language and decides which tools to call on each run. The agent tools docs describe three kinds: a single action from a connected app, one of your own flows handed over as a tool, or an external MCP server. You attach individual actions, so an agent given “Send message” can post to Slack and do nothing else there.

Between the two sits the AI Router. You write one question, each branch is a possible answer, and an evaluation model picks the route, with an Otherwise branch and an optional confidence floor. The docs advise the ordinary Router for numbers, dates and approval thresholds, because it is free and gives the same answer every time.

The limits page lists a 10-minute timeout per run and per action (time paused in Wait for Approval or Delay does not count), 1 GB of worker memory, 30 days for paused flows and run data, 10 MB per step file, 25 MB of combined run log, and 5 MB webhook payloads with a 30-second synchronous response. On your own servers each is an environment variable you can raise.

The crash recovery docs promise that completed steps never re-run after a crash or deploy, but the step that was executing does, so delivery is at-least-once. Make steps that send email or charge money safe to repeat.

The pricing FAQ says data is encrypted in transit and at rest and lives in the EU cloud in Frankfurt, another region on request, or your own servers. It also says the SOC 2 programme is in progress and offers a penetration test summary, DPA and sub-processor list on request. Self-hosted instances send product analytics unless you switch them off, and licensed instances send a daily snapshot of the deployment shape that the telemetry page says carries no flow data; it has its own switch. The Terms of Service page showed no document on the day we checked.

1. Sign up for the cloud Free plan, note the credit balance, and connect a test sheet and a test Slack channel.
2. Build a webhook trigger, an AI step on a fast model that labels a lead as sales, support or spam, a sheet row and a Slack post.
3. Send ten test leads, including two with an empty message, and open each run to check the step inputs and outputs.
4. Read the counter again and confirm it fell by about three credits a run.
5. Rename the sheet column for one run, let it fail, fix it and re-run, then confirm no lead produced two rows.

Accept if the counter matches the published rates, empty messages take the Otherwise branch and the retry creates no duplicate. Reject if credits fall faster than the rates explain, a lead is posted twice, or the flow needs a paid plan to finish. No result is recorded here.