An AI platform that runs on your own provider keys.
An OpenAI and Anthropic compatible model gateway, a chat that researches the web, and a Computer that builds apps and hosts them at their own address.
The Free plan includes the gateway, the chat and Context.
Any provider in the models.dev catalog, or a server of your own
- OpenAI
- Anthropic
- Mistral
- DeepSeek
- Groq
- xAI
- OpenRouter
- Your own Ollama
- Any OpenAI compatible server
How the products fit together
Sign in once. The gateway at api.nemu.cc, the chat at chat.nemu.cc and the Computer all use the same nemu account and the same session.
Metered and limited the same way. A conversation in the chat and a production request through the gateway count against the same workspace limits and appear in the same logs.
Context keeps notes your coding agents can read on any machine.
Short notes about each project: decisions, conventions, the gotchas that already cost time, and a handoff saying where the work stopped. The next session, on any machine and in any agent, starts where the last one ended.
$ nemu setup claude- Included on every plan, Free too, and it uses no credits
- Your own agent does the thinking. No model is called for Context
- A repository shares its notes on every machine you sign in from
Projects 3
- storefrontSharedgithub.com/acme/storefront42 notesOct 18, 2026
- billing-apigithub.com/acme/billing-api17 notesOct 16, 2026
- scratchNo remote6 notesOct 9, 2026
Notes 42
- Where the work stoppedhandoffCheckout tests pass, refunds nextOct 18, 2026Computer
- Sessions use rotating refresh tokensdecisionAccess tokens live for 15 minutesOct 18, 2026macbook-pro
- Webhook retries need the raw bodygotchaParsing it first breaks the signatureOct 18, 2026office-linux
Point the SDK you already use at api.nemu.cc.
Provider keys are sealed in the console. The gateway picks the model, enforces your limits, records the call and forwards it. Your code only ever holds a gateway key.
- Chat completions, the Responses API and Anthropic Messages at one base URL, translated in both directions
- Mappings give a model a stable name, so swapping what sits behind it never touches your code
- Your own Ollama or any OpenAI compatible server, at a URL the gateway can reach
- Rotating a provider key is one edit in the console, not a redeploy
curl https://api.nemu.cc/v1/chat/completions \
-H "Authorization: Bearer YOUR_NEMU_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "anthropic/fable-5.1",
"messages": [{ "role": "user", "content": "hello" }]
}'Requests are checked on the nearest gateway node.
The closest node checks the key, your quotas and rate limits, logs the call and forwards it to your provider. You can also choose a node yourself, on any plan.
Model Trigger
Metric
Model Selection
Model Selection
Choose the model at request time, based on latency, errors and spend.
Your client keeps sending one stable model name. A graph in the console checks latency, errors, spend and quotas, then picks the model and the server that handle the request. Your client code stays the same.
- Shed load onto another server when its average latency climbs
- Fall back to a second provider when a quota runs out
- Send requests that carry an image to a model that can see
- Serve a cheaper model at night or once spend crosses a line
What you pay for
Your provider bills you for tokens at its own prices, on your own key, and nemu adds nothing to that bill. What you pay us for is the infrastructure we run: the Computer, web search, page reads and deep search.
Who bills what
Billed by your provider
Model tokens
Requests run on your own key, so your provider charges its own rates and invoices you directly.
Metered from prepaid credits
Computer, search and page reads
- Computer memory
- €0.01 per GB-hour
- Computer processor
- €0.10 per vCPU-hour
- Processor, 00:00 to 08:00 UTC
- €0.075 per vCPU-hour
- Web search
- €0.006 a search
- Page read
- €0.003 a page
- Deep search
- €0.05 a run
- Indexing
- €0.20 per million tokens
Each plan sets how much of the gateway you can use.
Every plan uses the same gateway. Higher plans allow more requests a minute, more at once and larger bodies. A refused request says which limit it hit and when it resets. See every limit
- Free15
- Starter100
- ProPopular1,000
- Max2,500
- Enterprise10,000
Answers from the web link to every page they used.
The assistant runs searches, reads the pages behind the results and compares them before it answers. It works on the models you mapped in the console.
For one server, keep vectors next to the data. SQLite with a vector extension is a file beside the app sqlite.org, and Postgres with pgvector reuses the database you already run github.com.
A dedicated vector database only pays off once the set no longer fits on that machine.
- Web searchRanked results in one call, and the pages behind them read in the next.
- Deep researchOne run covers many searches and pages, then returns a written brief.
- SourcesOpen any source from the answer to check it yourself.
Chat jobs run on the server.
Use any model you mapped, and keep long answers as documents, pages or charts.
Runs in the background. Each thread runs as a job on the server, so research continues after you close the tab.
Any model you mapped. The model names you set in the console appear in the picker. Switch models mid thread without losing the conversation.
Artifacts. A long answer can become a document, a web page or a chart that you open, edit and keep.
Turn any web page into clean Markdown.
Send a URL and get the main content back. Navigation, cookie banners, ad slots and scripts are removed; headings, links and code blocks are kept.
Request
curl https://api.nemu.cc/reader \ -G --data-urlencode "url=https://example.com/blog/streaming" \ -H "Authorization: Bearer $NEMU_KEY"
- Cost
- 1 page read · €0.003
- Model in the loop
- None
Response200 · format: markdown
# Streaming responses Send stream: true and read the events as they arrive.Each chunk carries a delta. [API reference](https://example.com/api) ```pythonstream = client.chat(stream=True)```
- No model in the loop, so a page converts the same way every time
- PDF and DOCX come back as their plain text
- Drawn from the same credit balance, on every paid plan
The Computer runs what the chat builds.
Ask in a chat for something that needs to run. The agent creates the project, installs what it needs, builds it and serves it at a live URL. Nothing runs on your laptop.
bun, node, git and python3
A preview at its own address
Private, public or password
A 2 GB machine for every project
Project memory across chats
Cheaper from 00:00 to 08:00 UTC
You will learn about the Computer: what the sandbox can install and run, the 2 GB machine and how memory and processor are metered.
ComputerRates, limits and how a preview runs(opens the documentation in a new tab)You will learn about publishing: private, workspace, public or password, keeping a preview always awake and what a visitor can see.
PublishingWho can reach a preview(opens the documentation in a new tab)You will learn about credits: what draws from the balance, what never touches it and what happens when it runs out.
CreditsThe one balance it draws from(opens the documentation in a new tab)
Send your first request in about five minutes.
You need a nemu account and an API key from one provider.
- Step 01
Connect a provider
Pick one from the models.dev catalog or define it by hand, then add your own key.
- Step 02
Expose a model
Enable a model under a gateway name of your choosing, then create a key for your app.
- Step 03
Point your code at the gateway
Swap the base URL and the key. The rest of your code stays as it is.
Get started
The gateway, the chat and the Computer share one account, one set of API keys and one credit balance.