AI agent platform
The best agent for the job, at a cost you can see.
Build agents for any use case, ground them in your knowledge, and tune every setting. Lumagen scores each change on your own scenarios and prices every call, so better answers never come with a surprise bill.
What agents can do
One agent for support, sales, billing, onboarding, or the job nobody has named yet.
Build as many as you need. Each gets only what its job calls for: knowledge, tools, reasoning, a voice, the files people send.
search_knowledgeReturns policy · p. 7
create_returnRMA-4821 opened
Sorry about that. I've opened return RMA-4821 — a prepaid label is in your inbox.
Sources
Answer from your knowledge
PDFs, text, audio, video and images, cited down to the page, timestamp or region it came from.
Act through tools
Call HTTP APIs, MCP servers and web search, with credentials kept in project secrets, not prompts.
Think as much as it should
Full, brief or silent reasoning, the model's thinking level, and a budget of calls per reply.
Take voice calls
Its own prompt, first line, model and voice, cloned from a sample if you like, in 33 languages.
Read what people send
Images, audio, video, PDFs and documents, up to five files in a single message.
Knowledge
Drop in anything. Get answers that point to the exact page, second or pixel.
Documents, recordings, videos and images become knowledge your agents search — and every answer shows where it came from.
⌕refund window for opened items
Results
Returns policy.pdfp. 7
Opened items can be returned within 30 days if all accessories are included.
Anything goes in
Upload files, paste text or point at a link. Each source is summarized and split into chunks for you.
See how it connects
Every source becomes a graph of its chunks, so you can see what sits next to what before an agent ever reads it.
Tune what's retrieved
Embedding model, top_k, similarity threshold and surrounding chunks — then test a question and see exactly what comes back.
Every reply, itemized
Know what each answer cost, and why.
Every call is timed, counted and priced. Tune with numbers in front of you, not a bill at the end of the month.
Latency, taken apart
Dispatch, first token, thinking, tools and answer — see which part is slow.
Tokens by class
Cached, fresh, reasoning and answer tokens, with the cache-hit rate that keeps bills down.
Cost in your currency
Every call priced, every conversation totalled, in US dollars or any currency you pick at today's rate.
Usage over time
Cost, tokens and calls by model and agent, for the last day, week or quarter.
Latency
1.84s
First token
0.42s
Speed
86 tok/s
Calls
3
Cost
$0.00185
Where the time went
Tokens72% from cache
Where the money wentcache saved $0.00224
Context
≈16.9k / 1.05M · 2%
Cost · last 7 dayslargefast
Quality you can prove
Better answers, checked before anyone sees them.
Write the conversations that matter once. Every prompt, model or setting change is scored against them, so you know what improved and what it costs.
Scenarios with branching turns and more than one right answer
Scored on expected text, a weighted rubric with must-nots, and the tools it picked
A judge model you choose, repeated runs to smooth out noise
Optimization rewrites prompts and temperature until it clears your bar
temperature0.7→0.3
system prompt
search_knowledge· tool prompt
5.6× cheaperfor 2 points of accuracy. Try a model on a test run before you switch, and nothing is saved until you choose.
Try Lumagen
Build it, measure it, then ship it.
Start with a guided trial: we build your first agent with you, on your knowledge and your scenarios. Leave your email and we'll set it up.
A guided trial
We build your first agent with you, on your own knowledge and the conversations that matter to you.
A team alongside you
Work with us closely, from the first agent to every one after it.
Pricing you can negotiate
When the trial has shown what your agents can do, we'll agree on a plan that works for you.