Mole: A Deep-Research Agent That Actually Respects Your Budget

Every few weeks a new "deep research" tool shows up, and most of them share the same two flaws: you find out what the run cost after it's done, and the final report cites sources you just have to trust. Mole, a new open-source project that hit Hacker News this week, takes a different swing at both problems — and the approach is worth your attention even if you never install it.

What Mole does

Mole is a deep-research agent that lives in your terminal. You give it a question, and it decomposes the question, searches the web and academic sources, reads pages, extracts claims, checks each claim against the text it came from, hunts for contradictions between sources, and writes a cited answer. So far, so familiar.

Two things make it different.

First, the budget is enforced, not estimated. Every model call is reserved against a budget before it runs and settled after, with constraints built into the database schema itself. You pass --usd 0.50 and the run stops at fifty cents. The project reports zero measured overshoot across its test corpus. Anyone who has watched a research agent happily burn through API credits while "just checking one more source" knows why that matters.

Second, every claim carries a quote, and that quote must appear verbatim in the source it was mined from. If it doesn't, the claim is discarded before it can reach the answer. Claims that survive can be re-checked later, and any that turn out to be unsupported get flagged in the report rather than quietly dropped. It also queries Crossref, OpenAlex, arXiv and PubMed directly, deduplicates by DOI, and prefers open-access full text — so the citations are real papers, not blog posts quoting blog posts.

There's a privacy angle too: point it at a local CSV or folder and it analyses the data without the contents leaving your machine. Only aggregates — counts, means, test results, buckets covering at least five records — are allowed back to the model, and a mole crossings command shows you exactly what left. It speaks MCP, so a coding agent like Claude Code can drive it directly.

What it means for how you work

If you use research-heavy AI workflows — market analysis, fact-checking, literature reviews, due diligence — the interesting question isn't "does this replace my chat bot." It's whether a research run with hard budget ceilings and quote-verified citations is more trustworthy than a chat window with web search. For anything you'd put your name on, I'd argue it is. A hallucinated citation in an internal memo is embarrassing; in a client deliverable it's a liability.

The cost control story matters at the other end too. Deep research runs are exactly the kind of workload that scales silently — five dollars here, twenty there, until someone asks why the API bill tripled. A tool that makes the ceiling part of the contract, not a hope, changes the economics of letting agents run unsupervised.

My angle

I want to actually test this one. Research, fact-checking, and citation-based analysis look like the strong use cases — the architecture is built around verification rather than vibes, which is the right instinct. The honest open question is whether it significantly reduces hallucinations, and at what cost per answer compared to just asking a raw model. A 100% citation accuracy rate means the quotes are real; it doesn't automatically mean the synthesized conclusion is right, and that gap is where the real evaluation needs to happen. The project's own scorecard — 80% grounding rate on re-read claims, 70% contradiction precision — is refreshingly honest, and those are the numbers I'd want to reproduce myself.

One caveat on my side: it's written in Go, which isn't part of my stack, so if it needs patching I'm a user, not a contributor. That said, it ships as a single static binary with an Apache-2.0 license and works with any OpenAI-compatible endpoint — DeepSeek, Ollama, vLLM, whatever you're already running. Local models are priced at zero but still counted in tokens, so you can bound a self-hosted run that costs nothing at all.

That last detail is the quiet headline: budget-bounded, citation-checked research on top of a local model is a genuinely useful primitive. I expect to see this pattern — enforced budgets, verified quotes, local-data boundaries — copied into bigger platforms soon. Mole is worth watching precisely because it treats trust as an engineering constraint, not a marketing claim.