How Nexpher works

Nexpher turns plain descriptions into working AI assistants you can chat with, teach, share, and connect to your own product. This page explains every part of the service.

1. 01Create a free account — no credit card needed. The Free plan includes 50k credits a month and 200 MB of storage.
2. 02Open New agent and answer a few short questions — job, name, voice, apps. Each page helps the assistant take shape.
3. 03Try the draft before you save. After creating, chat with it immediately — those conversations can later power fine-tuning.
4. 04Keep it private, share it with the community (free or paid), or connect it to your own app with an API key.

The builder asks a few short questions — what it should help with, what to call it, how it should talk, and which apps it may use. We write the instructions. You can try the draft before creating it.

Prefer a head start? Templates give you pre-configured assistants (ops, research, support, writer, coder). Pick one from the homepage, tweak the prompt, make it yours.

After saving, add files for facts it should look up, or open the Train tab in Studio to teach it from chats you keep. Files help with facts; teaching helps with how it answers. They work together.

On Studio, Apps lets you sign in to Gmail, Calendar, Notion, Slack, and more. Tap Sign in and log in to that account — no API keys. Extra HTTP tools or custom servers stay under Advanced.

Knowledge is how you teach an agent facts, policies, product docs, or any reference material — without retraining a model. Upload .txt, .md, .csv, .json, or .pdf files (or paste text) on the Files tab in Studio. Add a website to crawl same-domain pages into a site tree. We chunk the content, embed it, and store it on the agent.

At chat time, only the relevant slices are retrieved (not the whole file) and treated as untrusted reference data — never as instructions. Replies show which files were used as sources.

Knowledge and fine-tuning are independent: documents do not become training data, and chatting does not update your knowledge base. Use both when you need grounded facts plus a tuned conversational style. Product API chats are stored when retention is on, then you choose which threads train the next run.

Every agent runs on a language model. Nexpher ships with fast open models out of the box — pick one from the model picker above the composer or set a default on the Model tab in Studio.

Each model burns credits based on its real cost, so heavier models consume your allowance faster. The picker shows which are enabled and their relative cost.

Switching models does not delete your agent. Knowledge docs, system prompt, tools, chat history, and saved fine-tune jobs all stay attached to the agent — only the active inference model changes.

Have an OpenAI-compatible endpoint — OpenRouter, Groq, Together, even Ollama on your own machine? Register it under Dashboard → Models with its base URL and API key.

Your endpoint appears in the model picker for your account only. Messages routed through it never touch our inference budget — we just relay them.

Studio is one workspace for the agent: profile, files, training, apps, and sharing. On Train you teach from chats: try the agent, keep the good replies, correct the bad ones, then train.

The loop is: Try in chat → if the answer is right, snap it into the lesson → if it is wrong, open the thread and tap Correct this (or Add lesson and type the reply you wanted). When you have enough, hit Train from this lesson, wait for succeeded, then Use this model. Repeat after you see new mistakes.

You can also paste or import user/assistant examples (JSON or JSONL) from Train → Import. Product chats from the IFS embed show up here when retention is on — Keep tab controls that.

Each user and assistant turn you keep becomes a lesson (at least 3, at most 2,000). Pictures in chat count toward storage and credits; teaching uses a short text note such as [1 image] instead of the file. Pick a model and start — we teach a small skills patch. The original model is not overwritten, and you can switch back anytime.

Keeping chats is separate from teaching: you can store a thread without using it as a lesson, or auto-delete product chats after a while.

A taught version is tied to the model you picked. Teaching on a new model reuses the chats you already selected.

Teaching from chats is included on Starter and above. Free accounts can build, chat, keep data, and add files.

Jobs go preparing → running → succeeded. Use a succeeded run to switch the agent to that version, or pick a standard model on the Model tab.

Share any assistant from Studio: open the Share tab, pick a category, and tap Share — free or paid per message. Earnings and payouts live under Share & earn in your account menu.

Other users install it and, for paid listings, pay per message from their wallets. Platform fee defaults to 0% (creators keep 100%); any fee is shown in-product and configurable by admins.

We'll email you when you've used about 80% and 100% of your monthly credits so you're not surprised.

Put your agent on any website with a floating chat bubble, sidebar panel, bottom bar, or inline iframe. Visitors chat without a Nexpher account — usage bills to you.

In Agent Studio → Website, preview the bubble on a light or dark page, set your logo (or keep the default Nexpher icon), colors, theme, and disclaimer, then create an embed key and copy the snippet. Turn on site actions so the agent can search and open pages on that site (same origin only).

Pass data-external-user-id (or external_user_id on the API) so each visitor gets their own monthly credit limit. Manage defaults and individuals from Dashboard → APIs → End-user limits.

Lock keys to specific domains (https://example.com) so stolen keys can't be used elsewhere. Control the widget from your page with NexpherEmbed.open(), .close(), and .toggle().

Every agent can get its own nxa_… key so your website, app, or backend talks to that agent directly. Point any OpenAI SDK at the agent base URL. Replies include conversation_id when you retain product chats — send it back to keep a support ticket on one thread for training.

Send external_user_id to meter each end user under a separate monthly credit budget (default 10,000 / month). Team agents work the same way: editors and admins can mint agent keys. Usage is billed according to the team's billing mode (owner, each member, or shared team plan).

Want Cursor or Claude Desktop to talk to your agents? Point an MCP client at POST /mcp with the same Bearer key — tools include list_agents, chat_with_agent, and get_credit_balance.

curl https://api.nexpher.ai/v1/agents/<AGENT_ID>/chat \
  -H "Authorization: Bearer nxa_..." \
  -H "Content-Type: application/json" \
  -d '{ "message": "Hi!", "external_user_id": "user_123" }'

Create a team, invite collaborators, and share agents under one workspace. Roles are owner, admin, editor, and viewer.

Pick how usage is billed: charge the team owner, each member individually, or a shared team plan with its own credit pool.

Plans bundle a monthly credit allowance plus storage. Credits are consumed by chatting, training and serving shared agents — cheaper models burn fewer credits than premium ones for the same reply.

Run low? Buy a credit pack (stacks until your period ends) or upgrade. Nothing auto-bills beyond your plan; agents pause politely if you hit zero rather than surprising you.

Prices and packs on the pricing page load live from our billing service, so what you see always matches what checkout charges — no stale screenshots.

Nexpher also has a desktop app for people who want a dedicated window instead of a browser tab. Mac ships first (macOS 12+). Windows and Ubuntu (.deb) are next. The Download page is the only place we will put real installer links — if a platform is not ready, the button stays Coming soon.

The app uses the same account as the website. Sign in, chat, open Studio to teach an assistant, and download models to run on your computer (photos and video keyframes included). Billing, API keys, community listings, and connecting Gmail or Calendar stay on the website for now.

Nexpher AI is operated by Nexpher Limited. Conversations, training files and trained adapters live in managed storage tied to your account. We do not sell personal data and do not use your content to train unrelated third-party models without your instruction.

Deleting a conversation, dataset or agent removes it from production storage. Full details are in our Privacy Policy, Terms of Service, Acceptable Use Policy and Cookie Policy.