Full report
The default. Built for decisions that need evidence.
- Planner chooses 1–6 research roles
- Fact-checker rejects claims, orders follow-ups
- Cover, contents, charts, diagrams, methodology
$ boku report "How are banks deploying GenAI?"
Open-source · Go · multi-agent
Research reports
Research reports
you can audit.
Give Boku a question. A team of agents plans the research, searches in parallel, lets an independent critic reject weak claims, and writes a typeset PDF where every sentence cites its source and every chart traces to evidence.
$go install github.com/riteshsonawane1372/boku/cmd/boku@latest
How it works
Ten roles, one evidence store, a critic with a veto.
This is a replay of a real run, compressed about 30×: 11 agent calls, 72 findings from 77 sources, two follow-up tasks requested by the fact-checker, $7.73 reported by Claude Code. Switch modes to see what changes.
Full mode timings and costs are the recorded values from run
2026-09-23T024958 (--depth quick).
Cost
What a report costs.
Spend is dominated by the research agents reading the web, not by writing. Numbers below are from real runs, as reported by Claude Code (total_cost_usd) and recorded in each run's manifest.json.
Recorded runs, by agent role
Each bar is one report; segments are the cost of each role.
Modes at a glance
Pick per report with a flag. Default is full.
| Mode | Typical spend | What you get |
|---|---|---|
| Fulldefault | $8–31measured | Planner, up to 6 researchers, fact-check rounds with follow-ups, 15–40 page report. |
Short--short | ≈ $4–6estimate | 3 researchers, one fact-check round, a compact 3–8 page brief. |
Quick--quick | $0local | Every agent on your Ollama model with Boku's own web search. No fact-check. A first pass. |
Whitepaperboku whitepaper | —new | Deep research, up to 6 researchers, two follow-up rounds. A 10–20+ page paper with abstract and references. Not yet measured; expect at least full-report cost. |
Explainerboku explain | —new | Up to 4 researchers, one follow-up round. A diagram-led explainer of a topic or a local codebase. Not yet measured. |
Short estimate: per-role costs of the --depth quick run above with 3 researchers and one fact-check round. Quick spends no API money; it uses your machine. On a Claude subscription, Claude Code usage counts against your plan rather than being billed at these figures.
Modes
One command. Five ways to run it.
No flag gives you the full report. Add one when you want something faster or cheaper, a whitepaper in research-paper form, or an explainer of a topic or a codebase. Combine with --save-ref to print the source list inside the PDF.
Short search report
A cited brief when you need the answer, not a book.
- 3 workstreams, quick depth
- One fact-check round, no follow-ups
- Compact layout: no cover, sections flow
$ boku report "Is Postgres 18 AIO worth it?" --short
Quick, on your machine
Every agent runs on a local Ollama model. Zero Claude tokens.
- Boku searches and fetches pages itself
- Models may only cite pages Boku fetched
- No fact-check: treat it as a first pass
$ boku report "GPU sharing on Kubernetes" --quick
Explainer new
Understand how a topic — or your codebase — works.
- Big picture, key ideas, glossary
- Architecture and step-by-step process diagrams
- Given a directory, agents read the repo (read-only) and cite files
$ boku explain ./my-service "auth flow"
Whitepaper new
The most detailed format, laid out like a conference paper.
- Abstract, keywords, numbered sections 1, 1.1, …
- Captioned figures and tables,
[n]citations, references - Deep research; results always carry their conditions
$ boku whitepaper "Sparse attention for long context"
Output
A typeset PDF, and a references file beside it.
Citations in the text are numbers. The sources behind them go to a compact <report>.references.json, cheap to hand to another model or tool. Pass --save-ref to print them in the PDF as well.
Pages from a real 32-page report, unedited. Open the full PDF.
kubernetes-platform-ai-infrastructure.references.jsonJSON
{"title":"Kubernetes as the Platform for AI Infrastructure",
"run_id":"2026-09-23T024958-kubernetes-used-ai-infrastructure",
"references":[
{"n":2,"title":"The CNCF Annual Cloud Native Survey: The Infrastructure of AI's Future",
"publisher":"CNCF","url":"https://www.cncf.io/reports/the-cncf-annual-cloud-native-survey/",
"published":"2026-01-01","accessed":"2026-09-23","tier":1},
{"n":5,"title":"Gartner Forecasts Worldwide AI-Optimized IaaS Spending to Grow 96% in 2026",
"publisher":"Gartner", …}
],
"evidence":[
{"id":"F052","claim":"Gartner forecasts worldwide AI-optimised IaaS spending will grow 96% in 2026 to about $42B…",
"status":"Estimated","as_of":"2026","refs":[5]}
]}
Web interface
Every option, in a browser.
boku ui opens a local web app, embedded in the binary. Start runs with any setting, watch stages, cost and the log live, then browse the plan, evidence and report. It shows runs started from the CLI too.
$boku ui
Bring your own model
Claude Code by default. Anything else with a config file.
Point Boku at Ollama or any OpenAI-compatible endpoint — vLLM, LM Studio, llama.cpp, OpenRouter. Nothing changes unless you pass a config.
- Boku's own web search (DuckDuckGo or SearXNG) for models without web tools
- Sources are re-grounded: a model cannot invent a URL
- Per-token prices in config give you cost tracking and
--max-cost - A local model fixes style problems so the main model isn't called again
config.yamlboku report "…" --config config.yaml
agents: provider: openai # claude-code | ollama | openai endpoint: http://localhost:8000/v1 model: Qwen/Qwen2.5-72B-Instruct api_key_env: OPENROUTER_API_KEY price_input_per_mtok: 0.35 price_output_per_mtok: 0.40 local: # small model for --quick and formatting model: llama3.1:8b search: engine: duckduckgo # or searxng max_pages: 12
Why trust it
Evidence first, prose second.
Claims keep their sources
Researchers return structured findings. The editor can cite only findings that survived fact-checking; citations are resolved by code.
Freshness is computed
Every finding carries the date it describes. Boku labels it current or historical against your window — the model never decides.
A critic with a veto
An independent fact-checker rejects unsupported claims and sends targeted follow-up research. What it still can't verify is published in a red Not verified box, never as fact.
Charts that trace
A chart is drawn only if every value appears in the cited evidence. Agents supply data, never layout.
Shaped by your question
A comparison opens with its matrix, a decision with the recommendation, a one-pager stays one page. Headings follow the request.
Inspectable and resumable
Every artifact lands in a run directory. Interrupted runs resume where they stopped; re-rendering costs nothing.