Mira is an integrated research environment for scholars working with primary and secondary sources across heterogeneous collections. It combines resource acquisition, archival management, full-text reading, model-assisted analysis, and structured composition within a single application — replacing the fragmented workflow of separate PDF viewers, browser tabs, reference managers, and word processors that characterises most contemporary research practice.
Most applications built on large language models are made to produce and organise text, whether drafting, summarising, answering questions, or holding a conversation, and they do that well. Mira is aimed at something qualitatively different. It began as a research method for testing theoretical claims in the social sciences by checking them, step by step, against the textual evidence available. For roughly its first two years this method existed only as programming code, with no clickable interface; the point-and-click version came later, to open the same approach to a wider circle of scholars in history, the humanities, and the social sciences.
Working this way is slow. A single prompt can take well over an hour to answer. Most of that time, roughly ninety per cent, goes to gathering and processing the relevant materials before any response is produced. The trade is depth. When the application is running well and used with care, Mira’s answers tend to be more comprehensive than those of a frontier model’s chatbot. This holds even on older or smaller language models. Mira does not match the powerful enterprise systems, but it can often, if slowly, do more than a chatbot alone.
Because Mira is open source and transparent, not only in its code but in much of what the interface makes visible, working with it carries a further benefit: over time it gives the researcher a concrete sense of how applications built on large language models actually operate. That same openness makes the system auditable, and it invites collective refinement, so the tool can be improved by the community that uses it.
Mira is a multifaceted web application with many parts and dependencies. This clickable diagram gives a general overview of how it runs: select any component to read, in plain language, what it does, and follow tool names to their own projects. The dashed line marks the researchers’ read-only access to your library — they can read it, never change it.
Much of the uncertainty around generative language models in the historical humanities and social sciences is blamed on a single flaw: hallucination, the tendency of a large language model to invent sources and produce plausible claims that turn out to have no basis in the historical or scientific record. Mira begins from a different diagnosis. The deeper issue is less hallucination in the abstract than a plausible-claim problem: an output can read as entirely credible while being wholly unsupported, and on its own a reader has no reliable way to tell the two apart. In other words, models do not ‘hallucinate’; they fabricate, and that process of fabrication is the same whether the output is subsequently judged ‘true’ or ‘false’.
Generating text with a language model is a fundamentally stochastic process, which is often taken to mean that these models carry an intrinsic potential to hallucinate. Developers narrow the distance between what a model produces and what a researcher wants by tuning models on subject-specific data and by retrieval-augmented generation. These measures help, but they do little to dissolve the scholarly uncertainty the plausible-claim problem creates: a fluent, well-formatted passage still cannot be trusted at sight.
Mira’s argument is that this uncertainty does not follow from the model alone. It is mediated by the application that hosts the model — the software that decides what the model is given, how its output is assembled, and how the finished text is presented to the reader. Once the plausible-claim problem and that surrounding application are placed at the centre, a language model can be used reliably for text-based research even if it remains, in itself, capable of fabrication. Mira does this with two instruments working together: a model-agnostic, graphically programmable multi-agent system that coordinates the models within a fixed programmatic structure; and a verbatim retrieval mechanism that extracts exact text from your source documents and stores each excerpt in a durable, machine-readable form the models can quote but never rewrite.
In practice this works like the copy-and-paste of an ordinary word processor. Text is copied verbatim from the source and pasted in beside the model-written prose; by interweaving exact quotations with generated text, Mira produces an auditable research essay that carries its own quote-verification and a standardised bibliography. The passages that reach the Writing Desk are identical to those extracted on acquisition, and each can be traced back to its origin in the source document.
Mira is not coupled to any single language model or commercial provider. The application supports whatever models the researcher has access to: small open-source models running on local hardware, cost-effective services such as Kimi & DeepSeek, providers like Mistral operating under EU data-protection regulations, or frontier models accessed through individual API credentials or subscriptions. We are also in the process of retraining our own, open-source, large-language model that is specifically catered to humanistic, historical, and social scientific research and operates in a transparent architecture.
A growing number of universities now host open-source language models on institutional infrastructure. Virginia Tech, for instance, operates over forty models through its Advanced Research Computing programme, available to students and faculty at no cost with no data transmitted to external parties. As institutional hosting becomes more prevalent, Mira is designed to integrate with these services as they emerge.
The application also employs machine learning to improve the effectiveness of smaller models — optimising not the model itself, but the architecture of its interaction with the researcher’s materials: context selection, query formulation, and result evaluation. A well-configured affordable model can, in many cases, approximate the performance of a considerably more expensive one.
Strictly speaking, Mira is entirely free. It is released under the GNU Affero General Public License (AGPLv3): anyone may run, study, distribute, and modify it, without payment or permission. What is not free is the electricity and hardware required to run it.
For researchers who cannot navigate the cumbersome process of setting the application up on their own hardware, an accessible version of Mira is made available on rented hardware. The primary way in is to apply as a researcher-collaborator: accepted scholars receive free access. Those who can't take part in development but want to use Mira now can take beta access on a non-profit basis — offered as-is while the program develops — which pays for convenience, not for the software, which remains free to anyone who wishes to self-host.
All revenues from beta access are publicly declared alongside our existing financial transparency statement. Mira is entirely volunteer-operated at the moment, and no one is taking a salary from the project. For the time being, all revenues are reinvested in full into the development, maintenance, and hosting of the application.
Development of a dedicated Mira language model is ongoing. Enabled by two small grants from Google, Mira has successfully performed its primary LLM-based computational functions on an open-source model — GLM Air, developed by Z.ai — running on a Google Cloud virtual machine. Carrying out this work on a fully open-source model, rather than a proprietary one, is a major step toward autonomous, community-based control of the underlying infrastructure.
The next step is to strengthen the infrastructure at the intersection of Mira’s multi-agent model harness and an open-source model, by fine-tuning the model in conjunction with that harness. Because these computational processes are entirely Mira’s own intellectual property and released under the AGPLv3 copyleft licence, their development and distribution are legally protected as freely and widely available, and subsequent work building on this system — already underway — is likewise protected for academic use. A fine-tuned Mira model is currently under development.
The present focus is building the computational infrastructure for fine-tuning open-source models. That infrastructure will allow the application to keep pace as new models are released, and to continually improve the model harness for research.
Mira is accepting applications from researcher-collaborators, and we help scholars apply for the Google Cloud research grants and other microgrants that fund the computing their work needs. Formal academic standing strengthens a grant application but is not a requirement; we consider every serious applicant on a case-by-case basis.
Mira is an open-source research platform operated on a full-disclosure basis. Its complete cost structure is published here and versioned in the public repository: every figure below is auditable, and every revision to it is part of the permanent record. Revenue will be published if any is received.
| Category | Detail | Amount |
|---|---|---|
| Hardware | Development machine | $400 |
| Engineering tooling | Agentic coding-tool subscriptions, 10 months | $1,340 |
| Cloud infrastructure | Compute, storage, and GPU sessions (Google Cloud) | ≈ $1,300 |
| Model usage | Development and evaluation (DeepSeek, OpenRouter) | ≈ $350 |
| Legacy hosting | Prior deployment during migration (Render), 12 months | $340 |
| Domain & DNS | Registration and name services (Cloudflare) | ≈ $50 |
| Total direct expenditure | ≈ $3,780 |
Figures marked ≈ are close estimates drawn from provider billing records. Contributed labor is not monetized as a line item.
The platform’s fixed running cost is a single small application server. The GPU host used for self-hosted open-source models remains dormant — it enters service only when revenue supports it — and costs nothing while stopped.
| Item | Monthly | Annual |
|---|---|---|
| Application server (24/7) | ≈ $98 | ≈ $1,175 |
| Storage | ≈ $15 | ≈ $180 |
| Network | ≈ $5 | ≈ $60 |
| GPU host (dormant) | $0 | $0 |
| Operating floor | ≈ $118 | ≈ $1,415 |
Model usage scales with actual demand and is additional to the floor; at zero usage it is $0.
Mira credits pay for the model tokens your work uses, itemized on every order before payment. Storage beyond the included allowance is $1 per GB per month. Hosted-access users receive technical support. Account creation is free, and non-model features carry no charge.
The entire work of making Mira run is volunteer-based. If you have technical skills and would like to volunteer, please contact [email protected].
Mira itself is free and open source; anyone may run it. What is offered for a fee is a ready-to-run copy, hosted and set up on rented infrastructure, so the fee pays for that convenience and not for the software. The primary way in is to apply as a researcher-collaborator: accepted scholars receive a ready-to-run Mira at no charge, our help securing a research grant, and a bridge to developer tools, in return for their disciplinary guidance. Everything that brings its own compute is free of the hosting fee: research cloud credits (GCP), self-hosting, and the developer add-on. Connecting your Claude Code, Codex, or OpenCode subscription is a cost-reducing add-on on top of any option, and it never adds a fee.
Annual hosted-access subscribers get priority access to Mira’s own fine-tuned open-source models as that capacity scales. Those models run and operate today, but not yet at the scale needed; Mira Methods is an academic nonprofit, and this revenue is what expands that capacity.
Accepted researcher-collaborators receive a ready-to-run Mira at no charge, with no hosting fee needed. How a GCP award is used, the balance between the recipient’s own research and improving Mira for the community, is decided with each recipient individually, based on their specific research and collaboration interests.
[email protected] will work with any interested PhD student or faculty member in the humanities and social sciences to secure the $1,000 and $5,000 compute credits under the GCP program — including writing the project proposal.
Prefer not to pay the margin? Mira also accepts a personal OpenRouter API key (in the app: Settings → Account & Credits) — usage then bills your own OpenRouter account directly, with no margin and on a self-serve basis.