Elements is a canvas for pipelines and agents built from models, your documents, APIs and code. Rehearse a run to see exactly what it would send, put a person in the loop where it matters, and deploy it as an API on your own servers.
Private beta. You need an invite code to sign up.
A rehearsal of the real “Support ticket triage” template, recorded from the engine. Branch decisions, skipped steps and “would send” previews are the product's output. The model replies and fetched data are stand-ins, because this page has no provider keys. Open this template in the studio
Templates
Six templates cover the jobs people ask about most. Each one comes with the settings it needs and its own test cases, and all six run end to end in our test suite on every change, with stand-in model replies. If one stops working, the build fails.
Support
A ticket arrives by webhook. It is classified, urgent ones alert the team, a reply is drafted, and nothing is sent to the customer until a person approves it.
Sales
A form submission arrives by webhook. It is scored against your ideal customer; strong leads go to sales in Slack and to your CRM, the rest are kept on a nurture list.
Operations
Turn the text of an invoice into structured fields. Extractions the model is unsure about, or that fail validation, go to a person before they are recorded.
Knowledge
Answers questions from your own documents and says where each answer came from. An agent searches the knowledge base as many times as it needs; prompt-injection attempts are stopped first.
Reporting
Every weekday morning, fetch a page or feed, summarise what matters, and email it. If the source is down the run stops there instead of emailing an empty digest.
Research
Give it a question. It reads public web pages, does the arithmetic with a calculator rather than in its head, and stops at a step and cost limit you set. Every step it took is kept as evidence.
Testing and release
A workflow that emails customers or writes to your CRM is hard to test by running it. Follow one real template, support triage, through each safety tool. Every step, message and hash below is the engine's recorded output.
Run the pipeline for real, except the steps that send, post or write outside it. Those report the exact message they would have sent, and the run carries on.
Each step that depends on an outside service is made to fail, once briefly and once completely. No provider is called, so a drill costs nothing.
A step that stops the run until a person approves, edits or rejects. They decide in the inbox, or from an emailed link without an account.
Change one step and run again from there. Earlier steps keep the results they already produced, so you do not pay for them twice.
Try a new version on real traffic. It runs beside the live one as a rehearsal, and callers only ever get the live answer. Then promote it or roll back.
An answer from a deployed pipeline can carry a signed receipt: which steps ran, which model, which path. Each step is chained to the one before; change anything and it no longer verifies.
Run the pipeline for real, except the steps that send, post or write outside it. Those report the exact message they would have sent, and the run carries on.
Each step that depends on an outside service is made to fail, once briefly and once completely. No provider is called, so a drill costs nothing.
A step that stops the run until a person approves, edits or rejects. They decide in the inbox, or from an emailed link without an account.
Change one step and run again from there. Earlier steps keep the results they already produced, so you do not pay for them twice.
Try a new version on real traffic. It runs beside the live one as a rehearsal, and callers only ever get the live answer. Then promote it or roll back.
An answer from a deployed pipeline can carry a signed receipt: which steps ran, which model, which path. Each step is chained to the one before; change anything and it no longer verifies.
Tests and a deploy gate. Save inputs with what must be true of the result. The triage template ships with “A billing problem with a deadline is urgent and waits for approval”. A test can check 12 kinds of things, and a deploy is refused if any test fails.
Branching
A decision step reads the input and picks a branch: a condition, a switch on a value, or a router that looks for keywords. Only that branch runs. The others are marked skipped, so you can see afterwards which path a run took and why.
I was charged twice this month.
After the first run
Run a pipeline over a list of inputs and fix single cells by hand. See what a change would do to last week's runs before you deploy it. Trace a sentence of an answer to the passage behind it. When a deployment keeps failing at one step, get a repair that was already tried on the failed calls.
| Input | Classify | Draft reply |
|---|---|---|
| I was charged twice this month | billing | Refund opened for the duplicate charge. |
| Invoice shows the wrong address | account | Updated the billing address on the invoice. |
| How do I reset my password? | account | Use “Forgot password” on the sign-in page. |
| Export fails with an error | bug | Thanks, the team is looking into the export. |
Four inputs run through the same pipeline, one row each.One cell edited by hand. Only the reply for that row ran again.
Before anything goes live
Rehearse any pipeline, put a person in front of what matters, and keep a record of every run.
Steps
A test fails the build if a step is listed in the palette without code behind it. The four GPU steps are off unless the server has a GPU: an administrator turns them on with one setting.
Email Inbox, Human Approval, IF / Condition, Only New Items, RSS Feed, Smart Router, Switch, Validator choose a path and only that path runs. Skipped steps are shown as skipped, on the canvas and in the run record. Independent branches run at the same time.
An agent step picks tools one at a time: web requests, your knowledge base, a calculator, sandboxed Python, your other pipelines, or any MCP server. It stops at the step and cost limit you set, and every step it took is kept.
Discord, GitHub, Google Sheets (add rows), Google Sheets (read), Notion, Send Email, Send Email (your mailbox), Slack, Telegram, plus any HTTP API. Webhook URLs are checked against the service they claim to be, so a step cannot be pointed at your internal network.
Agents can use tools from remote MCP servers, and every pipeline you deploy is also exposed as an MCP tool, so other assistants can call it with the same scoped API key.
Retrieval
Write a set of questions with known good answers once. Score each retrieval strategy on precision, recall, nDCG and faithfulness, and compare runs before you switch. The RAG step ships with 9 strategies, from basic to corrective and fusion.
| Strategy | Precision@5 | Recall@5 | nDCG@5 | Faithfulness |
|---|---|---|---|---|
| basic | 0.71 | 0.64 | 0.69 | 0.81 |
| multi_query | 0.78 | 0.74 | 0.77 | 0.84 |
| hyde | 0.75 | 0.70 | 0.74 | 0.82 |
| rerankingbest | 0.86 | 0.79 | 0.88 | 0.90 |
| fusion | 0.83 | 0.81 | 0.85 | 0.87 |
| corrective | 0.80 | 0.76 | 0.82 | 0.89 |
Documents
Upload any of 16 formats, watch it parse and split, and choose how it is chunked: 8 strategies, from fixed token windows to parent-child and code-aware. Collections are versioned, so an evaluation always compares like with like.
Team plans are billed per active seat with usage-based execution credits. Unused credits roll over for one billing cycle. Enterprise contracts can pool credits across workspaces…
chunk 12 · p.4 · parent_childBy the numbers
Every figure on this page is read from the product itself when the site is built, so it changes when the product does.
kinds of step for the canvas, in 8 groups
steps that send, post or write outside the pipeline, and every one of them stops at a rehearsal
kinds of check a pipeline test can make before a deploy is allowed
ready-made use cases; the line shows how many steps each one has (4 to 10)
Security
Secrets never enter a pipeline graph, workspaces never share state, and every deployed endpoint needs a key you can scope and revoke.
Provider keys live in an encrypted vault. Steps hold a reference like vault:llm, never the key itself, so an exported pipeline contains no secrets.
AES-256-GCM at restPipelines, collections, credentials and run history belong to one workspace. Retrieval in one workspace cannot reach another workspace's documents.
Checked on every requestDeployment keys are shown once, stored hashed, rate-limited per key, limited to the endpoints you choose, and can be revoked at any time.
SHA-256 hashedSelf-hosting
Runs you start in the studio stream step by step from the FastAPI engine. Queued runs and document ingestion go through RabbitMQ to workers, with Redis carrying progress and MongoDB holding state.
It runs the pipeline for real, including model and knowledge-base calls, except for steps that change something outside it: sending email, posting to Slack, Discord or Telegram, writing to Notion or GitHub, HTTP requests other than GET, and writes to the data store. Those steps report what they would have done and pass their input on. Model calls in a rehearsal cost what they normally cost.
Docker. The repository includes a production compose file that starts the API, a queue worker, the web app, MongoDB, Redis, RabbitMQ, ChromaDB, a reverse proxy that obtains TLS certificates, daily backups, and monitoring (Prometheus, Grafana, Loki). One script generates the secrets and another deploys and smoke-tests the stack.
Yes. You connect your own provider keys (OpenAI, Anthropic, Gemini, Groq, OpenRouter, and embedding providers). Keys are stored encrypted in a vault. Steps refer to them by name; saved, shared and deployed pipelines never contain the key itself, and it is decrypted only at the moment a step runs.
A pipeline follows the path you drew. An agent step decides for itself which tool to use next, within limits you set: a maximum number of steps and an optional cost limit. It can use web requests, your knowledge base, a calculator, sandboxed Python, your other deployed pipelines, and tools from remote MCP servers. Every step it took is recorded with the run.
It is a private beta: registration needs an invite code and things will change. What is on this page works today and is covered by automated tests. Things we have not verified are not claimed here; for example, provider calls in our test suite use stand-in replies, and the Google Sheets step can read but not write.
On your infrastructure. Documents, collections, pipelines, credentials and run history stay in the databases you run. Model calls go directly from your deployment to the providers you configure. Evidence receipts contain fingerprints of content, never the content.
Pipelines are stored as JSON graphs with version history, and you can export your account data at any time. Deployed endpoints are plain REST with API-key authentication, and are also available over the Model Context Protocol, so callers do not depend on an Elements SDK.
Free during the private beta. Bring an invite code and your own model keys, and keep your data on your own servers.