Description
TypeSafe for n8n
Ask a model that only judges: yes/no, pick-one and rate-it questions, each answered with a probability.
- What: an n8n node that asks TypeSafe’s Jev yes/no (Noul), pick-one (Choice) and rate-it (Score) questions about
- Why: a chat model writes prose you have to parse and guesses when unsure. This node returns a number for
- How: in n8n, Settings → Community Nodes → Install
@taifoon/n8n-nodes-typesafe, add your TypeSafe key, and ask
any item, and returns every answer with its probabilities on one output.
every answer, so your workflow can tell a confident answer from a guess.
“Is this a refund request?”.
Use it to classify tickets and rows, check an LLM’s input or output, or grade delivered work.
It is not for writing text, summarising, or reasoning about code.
item ──► TypeSafe ──► the item + answers (p, probabilities, confidence) + outcome: pass | fail | review
│
branch on it with n8n's own Switch or IF node, if you want to
Install
In n8n: Settings → Community Nodes → Install, then enter @taifoon/n8n-nodes-typesafe.
The package has no run-time dependencies; everything it needs ships inside the node.
Get a TypeSafe API key from console.typesafe.ai and create a
TypeSafe API credential. The test button makes one tiny real call, so you know at once whether the key works.
The node talks to TypeSafe directly with your key, and nobody else, including us, is in the path.
Running n8n for a team? Supply the key from a secrets file. Upgrading from 1.1 or
earlier, or from 2.x (three outputs became one in 3.0.0)? See Upgrading.
Try it in two minutes
1. Add a Manual Trigger and an Edit Fields node with a subject and a body: paste in any
support email.
2. Add TypeSafe, operation Ask. Leave Item to Judge alone; it already sends the whole item.
Add three questions:
– is_refund · Yes/no · Is the customer asking for money back?
– team · Pick one · Which team should handle this? · billing, technical, sales, abuse
– urgency · Rate it · How urgent is this? · None | Low | Medium | High
3. Run it. You get answers.is_refund.p, answers.team.value, answers.urgency.value.
4. Now label it. In Thresholds put {"team": {"minConfidence": 0.6}}. Tickets the model is sure about get
outcome: "pass"; the ones it is unsure about get outcome: "review". Add a Switch node on
{{ $json.outcome }} to send each kind where you want it.
Ready-made versions are in examples/. For the same questions on any System One host (TypeSafe, Venice
or another), use @taifoon/n8n-nodes-system-one.
Judge workflows, ready to import
Each one runs as imported on sample data, needs only your own TypeSafe key, and ends in a Switch on outcome with three
plainly named exits. Import
the file in n8n (Workflows → Import from File), pick your TypeSafe API credential, press Test workflow.
| # | Workflow | For | The judge decides | File |
|—|—|—|—|—|
| 1 | Grade an AI agent’s delivered work | teams running agents | did the delivery meet the task’s criteria? complete / reject / needs review | judge-agent-delivery |
| 2 | Grade a Base agent job from its on-chain record | agent marketplaces, ERC-8183 builders | the same, from the job’s public record, plus the unsigned evaluator call | judge-base-job |
| 3 | Fact-check a chatbot answer before it ships | support and RAG teams | does the answer claim anything its source does not say? send / hold / review | judge-answer-factcheck |
| 4 | Accept or return a freelancer’s deliverable | agencies, marketplaces | criteria met? accept / request changes / escalate | judge-freelancer-deliverable |
| 5 | Refund-dispute judge | e-commerce, marketplaces | refund / deny / a person, from the complaint and the evidence | judge-refund-dispute |
| 6 | QA an extraction against its source | finance operations | does every extracted value match the document? post / re-extract / review | judge-extraction-qa |
| 7 | Grade a batch of submissions | educators, bounty programmes | per row: verdict, probabilities and a receipt | judge-bulk-submissions |
| 8 | Check a translation keeps its meaning | localisation | meaning kept, nothing added or dropped? publish / send back / review | judge-translation |
| 9 | Moderate posts against your own rules | communities | which rule, if any, a post breaks: publish / remove / a moderator | judge-moderation |
| 10 | Triage support tickets | support teams | refund? which team? how urgent? | support-triage-direct |
Workflows 1, 2, 4 and 7 use Jev Options → Ask RUBRIC_v1: code checks the facts first, Jev answers four closed
questions, and the verdict comes with a receipt anyone can recompute. The others ask your own Noul questions and label
them with thresholds. Every file is tested end to end in test/templates.test.mjs, and imported and run in n8n 1.123.82.
Workflow 2 needs only your TypeSafe key: the evidence route is a public read. It outputs jev.verdict, the receipt and
the unsigned complete / reject call for the job’s evaluator seat. Nothing is signed or sent.
Three kinds of question
| You ask | You get back | Think of it as |
|—|—|—|
| Yes / no (“Noul”) | p, the probability the statement is true | a yes/no with a dial |
| Pick one (“Choice”) | the option, a probability for every option, and a confidence | a classifier that knows when it is guessing |
| Rate it (“Score”) | a level on a rubric you write, and a confidence | a ranking you can threshold |
Describe the options, and add an other, to sharpen a Choice. Ask every question that might matter in one node:
they are answered at once and independently, so ten cost about the same as one.
What you must know
outcome is pass only if every question in Thresholds passed. Add urgency too, and a good refund ticket “fails” for not being urgent.
review.legend.not calibrated on your data.
Cost: about 0.6 to 0.8 s and 450 input tokens for three questions. TypeSafe charges 0.042 USD per million input
tokens and nothing for output: roughly 0.00002 USD per item. Input is text or JSON; no images or audio.
More operations and options
TypeSafe’s answer untouched. Reply and Raw Output
Jev Options, and the standalone package @taifoon/jev.
Documentation
This package integrates one service: TypeSafe. It has no run-time dependencies, runs in n8n’s strict mode, is
published from GitHub Actions with an npm provenance statement, and every release must pass n8n’s
community-package scanner (npm run scan runs it locally on the source and the packed tarball).
Licence
Independent project. Jev and TypeSafe are products of TypeSafe AI, Inc., which does not endorse this package.
MIT. An independent community node by Taifoon. TypeSafe and Jev are
trademarks of TypeSafe AI; this project is not affiliated with or endorsed by TypeSafe AI.
