# promptev

> promptev: sovereign AI agents for business. Agents that answer from your company's data, see only what each person may see, and ask before they act.
>
> promptev is an AI agent platform for companies: build AI agents without code on
> your own documents, apps and data, and every person gets answers only from what
> they are allowed to open. Agents search your knowledge, use your tools (Google
> Workspace, Microsoft 365, Slack, Jira, databases, any API) and act, with
> approvals before risky actions, masking of sensitive data before a model sees
> it, and a full audit trail. Build in the console, with Ori (the builder
> assistant), or straight from Claude, ChatGPT and other AI apps over MCP.
>
> Runs on your own model keys (OpenAI, Anthropic, Gemini, DeepSeek and 34 more
> providers, including Azure OpenAI, Google Vertex AI, Amazon Bedrock and
> Cloudflare). promptev cloud by default (United States), your own storage and
> database if you want them, or your own cloud or premises on Enterprise.
> SOC 2 Type II. Pay as you go: 1,000 free credits for a 30-day trial, no card, no seats.
>
> Product: https://app.promptev.ai · Contact: team@promptev.ai

## Start here
- [Knowledge base](https://promptev.ai/promptev-knowledge-base.md): the full answer in one file: what promptev is, how it works, everything it connects to, how it is priced, how it compares, and the questions buyers ask. Fetch this first; the pages below are the same material split up.
- [Documentation](https://promptev.ai/documentation/): workspaces, projects and agents; knowledge and permissions; tools and approvals; orchestrations; the callable URL, REST API, SDKs and both MCP servers.
- [Pricing](https://promptev.ai/pricing/): credits, what each action costs, Enterprise.
- [Compare](https://promptev.ai/compare/): promptev beside Glean, Microsoft 365 Copilot, ChatGPT and Claude Enterprise, LangChain and LangGraph, n8n and Zapier, and building it yourself; each page says where the other product is the better choice.

## What promptev is, in one answer each
- **Who it is for:** companies that want AI agents their staff, customers and partners can actually use on company knowledge, without leaking what a person may not see. Common first agents: an HR policy assistant, an IT/support helpdesk, a sales or finance analyst over spreadsheets and databases, a legal contract reviewer, an engineering assistant over Jira, Confluence and Azure DevOps.
- **What makes it different:** permissions are enforced inside the search, per person, so an answer is built only from documents that person can open (following Google Drive, SharePoint, OneDrive and Dropbox sharing). Governance (approvals, masking, audit) is in every workspace, including the free one. You keep your model keys, and can keep your storage and database.
- **No code needed:** describe the agent's job in plain words, attach knowledge and tools, pick a model, choose who may use it, publish. Or ask Ori to build it for you.
- **Where agents answer:** the Agent library in promptev, Slack, WhatsApp, a chat widget on your website, the HTTP API, and any MCP client. The same agent, knowledge and approvals in every channel.

## Two products
- [All products](https://promptev.ai/products/): the two, side by side.
- [promptev App](https://promptev.ai/products/app/): the product a team opens: workspaces, projects, agents, orchestrations, knowledge with per-person permissions, connectors, channels, approvals, audit, credits.
- [Context Engine](https://promptev.ai/products/context-engine/): the permission-aware retrieval underneath, hosted by promptev. Developers use it directly through the Python and JavaScript SDKs, the HTTP API and MCP: ingestion (put documents in) and retrieval (search them).

## How it is organised
- **Workspace:** people, roles, billing, audit and where data lives. Roles: owner, admin, editor, viewer. Viewers only use agents (in the Agent library); editors build inside the projects they belong to; admins and owners manage members, billing and settings. Adding people is free.
- **Projects:** each has its own knowledge, tools, connectors, members, API keys and masking rules, and what one project connects is not reachable from another. Project roles are built permission by permission, and members can be added as a person or through a single sign-on group.
- **Privacy between people:** each person sees only their own conversations, whatever their role. Disabling a member removes their access at once and signs out the apps they connected; a removed member's projects pass to a workspace owner.

## Agents
- Built from plain-language instructions, knowledge, tools, a model and channels. Test the draft in the built-in chat (only people who can edit the agent can try a draft), publish a version, roll back to an earlier one. Each agent is set to answer strictly from its knowledge and tools, or flexibly with general knowledge added.
- **Models, on your own keys:** OpenAI, Anthropic, Google Gemini and DeepSeek built in; 34 ready-made providers (including Mistral, Groq, Together, Fireworks, Cerebras, OpenRouter, xAI, Meta, NVIDIA, Hugging Face, Baseten, Moonshot, MiniMax, Qwen, Z.AI and Vercel AI Gateway), most added with just a key and with their models kept in sync; or any OpenAI-, Anthropic- or Gemini-compatible endpoint of your own, with your own prices. Each agent can use a different model. The console warns before a model is retired and will not let an agent with tools run on a model that cannot call tools.
- **Cloud platforms:** five of the 34 are clouds, each signed in its own way: Azure OpenAI (your resource; models are your deployment names), Google Vertex AI (Gemini, and Claude on Vertex; a service account or an API key), Amazon Bedrock (Claude models; AWS region and access keys), Cloudflare Workers AI and Cloudflare AI Gateway.
- **A smaller model bill:** promptev agents do the same work for less on your own model keys; the dashboard shows what you paid and what promptev saved.
- **Built-in abilities:** search knowledge with citations; create files (Word, PDF, Excel, CSV, PowerPoint, HTML, Markdown, text, JSON); run code over spreadsheets for sums, pivots and charts (can be turned off per workspace); remember earlier conversations with the same person; ask the person a structured question form when it needs a decision; hand parts of a long job to helper agents that share a work board; keep versioned artifacts it can update later; know the person's time zone.

## Knowledge and permission-aware answers
- [Permission-aware answers](https://promptev.ai/features/permission-aware-answers/): permission-aware AI agents and permission-aware RAG.
- Upload files (PDF, Word, PowerPoint, Excel and CSV, images and scans, e-mail files; 25 MB each, 20 per upload) or connect a source and pick exactly which folders, files, spaces or tables it reads. Sources stay in sync, with progress per file.
- Follows Google Drive, SharePoint, OneDrive and Dropbox sharing; uploaded documents get audiences. The person's identity and groups are applied during the search, so a document they cannot open is never fetched for them. Answers cite their sources. Hybrid search, plus an optional Context Graph of entities and relationships, kept with the documents (no separate graph database).

## Connectors and reach
Full list: https://promptev.ai/integrations/
- **Eight apps, 16 services, 198 ready-made tools:** Google Workspace (Drive, Docs, Sheets, Slides, Gmail, Calendar), Microsoft 365 (SharePoint, OneDrive, Outlook, Outlook Calendar), Dropbox, Confluence, Jira, Azure DevOps, Slack and WhatsApp.
- **Follow source permissions:** Google Drive, SharePoint, OneDrive and Dropbox.
- **Connect on promptev's apps or your own:** start with promptev's ready connector apps, or register your own so consent and quotas sit in your own tenant. Credentials are set up in promptev itself and are never shown to an AI.
- **Three open doors for everything else:** any REST or HTTP API becomes a tool; databases (PostgreSQL, MySQL, MariaDB, Microsoft SQL Server, Oracle, ClickHouse, Snowflake) answer plain-language questions with tables or charts, read-only by default or writable, limited to the tables you pick (a read-only tool runs exactly one reading query at a time, and anything that would change data is refused before it reaches the database); and any remote MCP server connects by URL. Every one of them is governed like a built-in tool.

## Approvals: people stay in charge of actions
- Each tool has a rule: **no approval** (right for reading), **ask once** (the first use in a conversation waits for a person, then that conversation carries on for the destination that was approved; a new recipient, channel or address asks again), or **always ask**. Tools you register yourself (HTTP, database, MCP) can also **ask only if** a condition holds, such as an amount over 1,000. An admin's rule on a tool is a floor an agent can tighten but never weaken, and a refusal never becomes a standing yes.
- **Who decides:** the person in the conversation or a workspace admin in the App; the person who asked, in Slack; or a named approver by e-mail (never the person who asked). In AI apps that can show it, the approval appears right in the app.
- Requests expire (60 minutes by default) instead of waiting for ever.

## The chat experience
- A live line shows what the agent is doing, then "Used N tools" with every step behind it. Citations, attachments (files, images, voice notes), previews of Word, Excel and PowerPoint, and charts (bar, line, area, pie, scatter) in answers.
- Question forms when an agent needs something from you: several questions become a step form, with text, e-mail, URL, date, number and yes/no answers.
- Copy, retry, stop, and thumbs up or down with a reason on every answer, so teams can see how their agents are doing. A turn keeps going if you leave, and is there when you come back.
- **Artifacts:** every file an agent makes is versioned and kept on the Artifacts page (yours, and shared with you), shareable with named people or by a view-only link (an admin can switch public links off), and can be handed back to an agent to change.
- **Agent library:** where people find and use the agents they are allowed to.
- **Usage dashboard:** model spend, what promptev saved and wait times, by provider, model and agent.

## Orchestrations and triggers
- An orchestration is a lead agent directing a team of agents. Start it by hand from an intake form (including files), on a schedule, from a callable URL, or from another agent. Runs are durable with retries, can pause for a reviewer's sign-off or be sent back for changes, and show every step and its cost.
- Agents can also run on a schedule or on an app event: an e-mail arriving in Gmail or Outlook, a Google or Outlook calendar event changing, a Jira issue updating, or a Confluence page updating.

## Channels
- [Channels](https://promptev.ai/features/channels/): one published agent everywhere.
- **Agent library:** signed-in members, each with their own access.
- **Slack:** mention it in a channel (it replies in the thread) or send a direct message. People matched to members by their Slack e-mail get their own access; everyone else gets public knowledge. Question forms, approval buttons, charts and files work in Slack.
- **WhatsApp:** public knowledge; documents, images and voice notes in, files back out.
- **Website widget:** one script tag, rendered in your page, only on the sites you allow; public knowledge.
- **HTTP API and MCP:** see below.

## Build from your AI app: the workspace MCP
promptev is an MCP server you sign in to. Add `https://api.promptev.ai/mcp` to Claude, Claude Code, ChatGPT, Cursor, GitHub Copilot, Windsurf, Gemini or any MCP client that supports remote servers, and sign in with your promptev account. No key to copy.
- On the sign-in page you pick one workspace and choose what the app may do: see projects, models and connections; see and edit agents; attach knowledge and tools; upload documents; publish; ask your agents; run your projects' tools; start connecting apps, provider keys and tool credentials.
- The app then works as you, with your own role and project permissions, never more. It can list projects, models, connectors, knowledge and tools, create and edit draft agents, set their knowledge and tools, upload documents, add knowledge from a connected app, and publish. It can also use what you built: search knowledge, find and run tools, call agents, and list, read, create and update artifacts, in any project you can use, under the same approvals.
- Creating a project and publishing an agent ask you to confirm first, in your AI app.
- **Secure links for steps only you may do:** connecting an app, adding a provider key, setting a tool's credentials, uploading files or setting up a channel are finished by you on a promptev page, through a link that works only for you, must be opened within 5 minutes and is good for 30 minutes. The AI app never sees a secret.
- **Built-in guides** (building an agent, fixing one, knowledge, models, connecting apps, tools and approvals, channels, permissions) tell your AI app how promptev works.
- **Feedback to promptev:** both MCPs carry a `send_feedback` tool (blocker, failure, improvement, suggestion, praise, other, with a short message), so an agent or app can tell promptev's team what got in the way. Free; promptev records which workspace, project and app it came from.
- Builders only (editors, admins, owners). Apps may connect to every workspace unless an admin turns off "Allow apps to connect" in Settings; everyone sees and signs out their own apps under Profile, in "Where you're signed in", and removing or demoting a member signs their apps out. Every change is in the audit trail under the person's name, with the app that made it.

## Ori, the builder assistant
Ori is promptev's friendly 3D guide in the console. Ask in plain words ("build an agent that answers HR questions from our handbook") and Ori does the work with the same tools as the workspace MCP, acting as you with your permissions: it knows the page you have open, looks at what you have, asks you when it needs a decision, hands you a secure link instead of ever taking a password or key in the chat, creates and configures the agent, and asks you to confirm before creating a project or publishing. Every step it takes is shown, your conversations with Ori are saved privately (earlier ones in a dropdown in its panel, which you can drag wider), and you can rate each answer. You can attach files to a message to Ori (the same types the agent chat takes), it can add knowledge from an app you connected, and an approval you may give appears as a Yes or No card in its chat. For editors, admins and owners; runs on promptev's own model at no extra charge at launch.

## From your own code: callable URL, REST API, SDKs, project MCP
- **Callable URL:** every agent and orchestration has one. An agent answers in the response; an orchestration is accepted and returns a run id to follow. The JSON keys are the orchestration's declared inputs, and a missing required one is refused with a 422 naming the field. Files go as multipart (a `payload` part plus one part per file, named after the input it fills; five at most). Retries are safe with an Idempotency-Key.
- **REST API with scoped keys:** a key belongs to one project, reaches that project's published agents and orchestrations and nothing else, and can expire or be revoked. Keep it on your server.
- **Options on one call:** a call to an agent may name a `model` (another model of a provider the workspace has connected, on the workspace's own key) and `tool_inputs` (replace a header, query parameter or body value one of the agent's tools is set up with, addressed by the tool's ID, so the agent can call your API as your end user). Nothing a call sends is stored. In the REST API, and in both clients (Python and JavaScript / TypeScript) from 1.1.0.
- **Files with a message:** a chat call to an agent can carry up to 5 files (25 MB each) in the same request, as multipart with the JSON in a `payload` part. Accepted only when the agent's Chat attachments setting is on (409 otherwise) and only the types the chat takes (415 otherwise); an image is shown to a model that can see. In the REST API, and in both clients from 1.2.0 (`files`).
- **Two official clients:**
    pip install promptev            →  from promptev import PromptevClient
    npm install @promptev/client    →  import { PromptevClient } from "@promptev/client"
- **Project MCP:** each project is also an MCP server at `https://api.promptev.ai/api/v1/mcp/<workspace>/<project>/mcp`, authorised with a project key, for tools and services: list and call agents, search knowledge, find and run tools, upload documents, and list, read, create and update artifacts, under the same approvals and audit.

## Governance and security
- [Governance](https://promptev.ai/features/governance/)
- **Masking:** e-mails, phone numbers, US social security numbers, card numbers, IBANs and API keys are caught, plus your own patterns; masked, hashed or removed in documents and tool results before a model reads them, with the real value shown only to the people you name.
- **Audit trail:** every run, approval and settings change, with who did it.
- **Single sign-on, in every workspace:** OpenID Connect and SAML 2.0 (Microsoft Entra ID, Okta, Google Workspace, ADFS and others), side by side if you are moving providers. Company domains proven by a DNS record send people straight to your identity provider from the sign-in page, and decide who may join automatically. SCIM 2.0 provisioning adds, updates and removes people and groups. An owner can send the IT admin a setup link (1 to 14 days) that needs no promptev account, and Test connection shows what the provider sends before anyone relies on it. SSO can be required for everyone, or everyone except viewers; SSO groups become access rules and can set roles.
- **Sign-ins and devices:** sign in with an e-mailed link or SSO, no password. "Where you're signed in" on the profile lists every browser, device and connected AI app, each of which can be signed out. A web sign-in ends after 48 hours without activity; every sign-in has a hard limit (30 days for a link, 24 hours for SSO).
- **Your data:** never used to train models; encrypted at rest (AES-256); conversation files kept 30, 60, 180 or 365 days, or for ever in your own bucket.
- **Moving to your own database:** when a workspace connects its own PostgreSQL, its existing documents move into it automatically (copied as they are when the embedding model is the same, embedded again when it differs), with a "Migration in progress" notice, a notification to owners and admins when it is done, and a retry for anything left behind. Its tools, their approvals and their call records move there too; tool credentials stay encrypted with promptev's key.
- **Bring your own storage** (Amazon S3 or compatible, Google Cloud Storage, Azure Blob) and **your own database** (your PostgreSQL, your embedding model on your own key, the Context Graph stored in that same database, and optionally your own reranker: Cohere, Voyage, Jina or any Cohere-compatible endpoint).
- **Trust:** SOC 2 Type II; promptev cloud runs in the United States; DPA, SLA, sub-processor list and a status page (status.promptev.ai).

## Pricing
- [Pricing](https://promptev.ai/pricing/): pay as you go at $0.005 a credit. $5 minimum top-up (1,000 credits). 1,000 free credits, given once at sign-up for a 30-day trial, no card; free credits not used in 30 days expire, credits you buy never expire. No seats and no plans: every feature is in every workspace, and model usage is on your own keys.
- **Lower AI cost:** in promptev's own tests on large data (October 2026), agents answered at up to 83% lower cost and with up to 78% fewer tokens than before that release; the saving is largest on questions that need a lot of data. Details: https://promptev.ai/ai-agents-83-percent-lower-cost-large-results/
- **What a credit buys:** a chat turn 1; a tool call 1 (including every tool call a connected app makes through the workspace MCP, building or using; a refused call costs nothing; working with Ori in the console is not charged per call); a triggered run 1; an API call 1; a knowledge search +1 (+5 when the Context Graph is searched); running code +1.
- **Reading documents in:** 1 credit per page of a PDF or Word file, per slide or per image, and per MB for spreadsheets and other files (minimum 1); a voice note 2. Embedding is free.
- Auto top-up, receipts and a payments ledger; low-credit notices at 80%, 90% and 100% of your last top-up.
- **Enterprise:** an annual licence for a deployment in your own cloud (any region) or on your own premises, with onboarding and support. It buys a place to run and people to call, not features. Sized to the estate, not to seats, and the route for data that must stay in one jurisdiction. team@promptev.ai

## Solutions by team
- [All solutions](https://promptev.ai/solutions/)
- [Legal](https://promptev.ai/solutions/legal/), [Finance](https://promptev.ai/solutions/finance/), [Engineering](https://promptev.ai/solutions/engineering/), [Support](https://promptev.ai/solutions/support/), [HR](https://promptev.ai/solutions/hr/), [Sales](https://promptev.ai/solutions/sales/), [Education](https://promptev.ai/solutions/education/)

## How promptev compares
- It does not replace the assistant your company already uses (Microsoft Copilot, Glean, ChatGPT Enterprise). Those are good at their jobs, and several are also permission-aware. promptev is for the agents you build yourself: your instructions, your tools, your approvals, on your model keys, reachable from Slack, WhatsApp, your website, your code and your AI apps, priced by use rather than per seat, and able to run in your own cloud.
- Compared with building on an agent framework, promptev gives you permission-aware retrieval, connectors, approvals, masking, audit, channels and billing ready-made, so a team ships an agent instead of a platform.

## Context Engine (hosted, through the SDKs, the API and MCP)
> The retrieval and permission layer underneath promptev. promptev runs it; it is not open source and there is no engine to install. You use it through promptev's SDKs and API: ingestion (put documents in) and retrieval (search them). Each search returns only what the caller may see (a project API key reads its own project and nothing outside it; people in the app see only what their access allows), and every hit names its source.
>
>   pip install promptev            (Python 3.8+)   from promptev import PromptevClient
>   npm install @promptev/client    (anywhere with fetch)   import { PromptevClient } from "@promptev/client"
>
> Python: `client.upload_document("handbook.pdf")` or `client.upload_documents([...])` (up to 20 files, 25 MB each; `mode="graph"` also builds the Context Graph of entities and relationships), then `client.search_knowledge_base("search", query="how many days do customers have to ask for a refund", top_k=5)`, which returns `hits` with document_id, document_name, chunk_text, score, source_id and chunk_idx. Async: `aupload_document`, `aupload_documents`, `asearch_knowledge_base`. JavaScript: `client.uploadDocument(files, { mode })` and `client.searchKnowledgeBase("search", { query, top_k })`.
>
> HTTP (base https://api.promptev.ai, header `Authorization: Bearer fk_...`, a project API key made in the promptev console): `POST /api/v1/documents` (multipart `files`, optional `mode`) and `POST /api/v1/knowledge` (JSON such as `{"action": "search", "query": "how many days do customers have to ask for a refund", "top_k": 5}`).
>
> MCP: every project is an MCP server; its `search_knowledge_base` tool is the same tool with the same actions, and `upload_document` adds a file. Actions: discover, search, get_doc, list, query_meta, compute, get_chunks, get_docs, map_reduce, traverse, find_related, get_neighbors, community_summary.
>
> Cost (credits at $0.005): ingestion 1 per page, slide or MB (minimum 1); a knowledge call 1, plus 1 for a search, or 5 when the Context Graph is searched. Through the SDKs and the API there is no extra API-call charge for knowledge calls. The rates are the same for a workspace on its own database or models.
>
> Bring your own: a workspace can keep its knowledge in its own database and use its own embedding model, its own OCR (vision) model and its own extraction model for the Context Graph, on its own keys.

- [Context Engine product page](https://promptev.ai/products/context-engine/)
- [How the promptev Context Engine works](https://promptev.ai/context-engine-open-source-retrieval-library/): on 100,000 documents with synthetic team-shaped permissions, a caller who may see 10% or 1% of the corpus gets the same top ten as a corpus holding only their documents (recall is agreement with that corpus, not relevance), while filtering after the search returns nothing for 63.5% and 99.0% of hybrid searches and recovers 0.638 and 0.083 of the results even when it fetches fifty times more first.
- [Your ACL benchmark is measuring the easy case](https://promptev.ai/acl-benchmarks-measure-the-easy-case/): the benchmarks of access-filtered search we know of assign permissions at random, but real permissions are clustered by topic. On 100,000 documents with synthetic permissions, for a caller who sees 10%, a hybrid search filtered after a tenfold fetch recovered 0.730 of the results with random permissions and 0.419 with clustered ones, with 13.0% of searches returning nothing. Why, and how promptev avoids it.
- [EU AI Act Article 12: what your RAG system has to log](https://promptev.ai/eu-ai-act-article-12-rag-logging/): what Article 12 asks of high-risk systems, how the Digital Omnibus deferred those obligations, and what an audit trail does and does not cover.
- [The Supabase extension trap](https://promptev.ai/supabase-pgvector-search-path-trap/): why `CREATE EXTENSION pg_trgm` succeeds on Supabase while the functions stay unreachable, and the one-line fix.

## Blog
- [Blog](https://promptev.ai/blog/): context engineering, permission-aware RAG, AI agents, MCP.

## Legal
- [Privacy](https://promptev.ai/privacy-policy/), [Terms](https://promptev.ai/terms-of-service/), [DPA](https://promptev.ai/dpa/)

## Everything else
- [Sitemap](https://promptev.ai/sitemap/) · [XML sitemap](https://promptev.ai/sitemap-index.xml)