betaFeedback

WorksComputer Science

ProofStudio: Five Proof Assistants in the Browser

Abstract

Recent advances in AI-assisted proof automation are extending the use of proof assistants in software verification and mathematics. However, greater proof automation does not remove users' responsibility to assess whether formal statements capture their intended requirements, making education in formal specification and reasoning essential. To lower the installation and learning barriers associated with mature proof assistants, we present ProofStudio, a browser-based environment integrating Isabelle/HOL, Lean, Rocq, HOL4, and HOL Light with corresponding worked examples, preconfigured proof search, and contextual AI assistance. With ProofStudio, users can explore formal specifications and proofs across five systems without local installation, inspecting native prover feedback and experimenting with automation through a common interface.

Introduction

Recent advances in AI-assisted proof automation are expanding the practical scope of formal verification. Anthropic and Theorem recently reported using Claude to produce Rocq-checked proofs of no-crash properties for 239 entry points across 123 open-source libraries, subject to explicit execution assumptions and input bounds [1]. In mathematics, OpenAI released a large collection of AI-generated manuscripts, many accompanied by Lean formalizations [2]. These developments make proof assistants increasingly relevant to software engineers and scientists beyond the formal methods community.

Greater automation, however, does not remove the need to understand and scrutinize what is being proved. OpenAI withdrew three manuscripts after a sign error invalidated an argument and two dependent proofs [3]. Beyond errors in reasoning, a separate concern remains even when a formal proof is accepted: verification establishes that a statement follows from its assumptions, whereas validation asks whether that statement captures the intended requirement. Anthropic’s report explicitly identifies the adequacy of specifications and machine models as limitations of its guarantees [1]. Recent work on AI-generated mathematics likewise reports discrepancies between natural-language arguments and their formalizations [4]. Scientists and engineers therefore still need the skills to read formal statements, inspect assumptions, and assess the scope of automated results.

Developing these skills requires opportunities to experiment. Mature proof assistants offer different logics, libraries, proof languages, and automation, with complementary strengths. However, trying each system often requires a separate installation and editor configuration. These barriers affect newcomers and make it harder, even for experienced users, to explore other provers. Corresponding worked examples can help users compare how different systems specify and verify familiar algorithms.

Local experimentation also requires trust in the installed software. Earlier Lean 4 installation instructions explicitly acknowledged this requirement [5], while the current guide includes a publisher-trust prompt for the Lean extension [6]. Lean’s official FAQ further warns that opening files, building projects, or updating dependencies can execute arbitrary code, and recommends isolation for potentially malicious code [7]. Remote hosting allows initial experimentation without installing or executing prover software on users’ own machines, shifting responsibility for installation and execution containment to the service.

We present ProofStudio, a browser-based environment for exploring Isabelle/HOL, Lean, Rocq, HOL4, and HOL Light. Its contribution is the integration of:

  • installation-free access to five native proof assistants through a common interface;

  • ten corresponding examples per prover;

  • preconfigured proof-search tools for supported provers; and

  • contextual AI assistance with a choice of model provider.

Users edit native proof source, inspect intermediate proof states, and review proposed source changes before accepting them. The intended users are learners exploring formal reasoning and experienced prover users trying another system.

User Experience

Figure 1 shows the Source, Output, and AI Assistant panes. The menu selects the prover, example, and search tool. At https://proofstudio.org/, information buttons beside Proof search and Output provide usage guidance.

ProofStudio with the Isabelle insertion-sort example

Figure 1. ProofStudio with the Isabelle insertion-sort example: ① prover and example selection; ② the Source pane; ③ the Output pane, showing intermediate subgoals and a successful whole-document check; ④ the AI Assistant pane, explaining why correctness requires both sortedness and count preservation. Public-service capture, 21 October 2026.

Editing and Native Prover Feedback

The Source and Output panes support the basic cycle of proof development: edit a definition or proof in the Source pane, then inspect the prover’s response in the Output pane. The Source pane supports basic syntax highlighting and the Unicode symbols used in the five proof assistants.

Persistent prover sessions provide automatic feedback during editing, while the caret position identifies the part of the proof the user wants to inspect. In Figure 1, for example, the Output pane shows the two subgoals after applying the induct tactic, because the caret is positioned immediately after induction on xs. These describe the local proof state at that point. The subsequent commands shown in Figure 1 complete the proof.

Checking Documents and Searching for Proofs

The Source pane offers two explicit commands: Check and Prove. Check requests native diagnostics and proof information, including feedback on unfinished documents. Successful processing is distinct from proof completion: if remaining goals are admitted using commands such as sorry in Lean and Isabelle or Admitted in Rocq, Check accepts the document even when proofs are still incomplete.

When the Source pane contains such holes, users can try the Prove button, which searches for a proof script that fills a proof hole. Table I summarizes the available search options for Prove, which users can choose in the main menu bar at the top. These include AbductionProver [8] for Isabelle, Aesop [9] for Lean, CoqHammer [10] for Rocq, and TacticToe [11] for HOL4. The current Prove command uses conventional proof search; LLM-based assistance is provided separately through the assistant pane.

ProverLibrary scopeProve options
Isabelle/HOLMainAbductionProver, Sledgehammer, try0
Lean 4CorePortfolio, Aesop, grind
RocqStdlibCoqHammer
HOL4CoreTacticToe, standard automation
HOL LightCoreNot supported

Table I. DEFAULT OFFERING IN V2026-10-04.1. ALL FIVE SUPPORT CHECK. LIBRARY SCOPE EXCLUDES SEPARATELY BUNDLED SEARCH TOOLS.

Upon successful proof search, the Prove command checks that source outside the hole is unchanged and uses a separate verifier to replay and audit the target theorem, as shown in Figure 2 in Section III.

Logical execution paths showing the browser user interface, manager, model provider, interaction, search, and verification

Figure 2. Logical execution paths; workers and return paths are omitted. Results return through the manager. AI edits require user review; repairs produced by Prove undergo source-preservation checks and prover-specific verification. Guest containment has the limits discussed in Section III-A.

Corresponding Examples Across Five Provers

ProofStudio provides ten corresponding examples for each prover: insertion sort, merging, ordered-list sets, binary search trees, two-list queues, binary tries, leftist heaps, binomial-tree linking, Braun trees, and Fibonacci pairs. Adapted from Functional Data Structures and Algorithms: A Proof Assistant Approach [12], these self-contained developments let users compare native specifications and proof techniques without configuring additional libraries. For the insertion-sort example in Figure 1, users can keep sortedness and preservation of element counts as the intended requirements while exploring how each prover expresses and establishes them.

AI Assistance and Provider Choice

All five provers share an AI Assistant interface with two modes: the document context mode and question only mode. In document context mode, the manager combines the current source, available goals, caret position, output, and conversation history with prover-specific instructions. This is the mode shown in Figure 1, where the assistant explains the two parts of the sorting contract. In question only mode, it omits document context and disables AI-generated source edits.

Users can choose the shared default LLM or an external service through bring-your-own-key (BYOK) configuration. BYOK requires an API key and a public HTTPS endpoint supporting the OpenAI-compatible interface. Section III-B explains the associated credential handling.

The model has no direct access to shell or prover tools. Proposed source edits are held for user review, and accepting a proposal does not certify its correctness. Users must inspect any changed specification and check the result with the prover.

System Architecture

Figure 2 presents the architecture of ProofStudio. The manager coordinates native prover interaction, automatic proof search, independent checking of search results, and access to AI model providers.

Prover Containment and Limitations

Processing user-supplied proof documents can execute potentially malicious code. Lean’s FAQ warns about this risk [7]. Remote hosting therefore requires containment of potentially hostile computations.

ProofStudio uses QEMU/KVM virtual machines, which provide hardware-assisted virtualization on Linux. Each proof assistant instance runs inside a virtual machine, while the manager serves the browser interface and schedules workers that control prover execution.

Prover VMs have no virtual network interfaces. During interactive proof development, workers exchange prover input and output through a dedicated communication channel provided by QEMU. Users interact with the provers through ProofStudio’s web interface. ProofStudio also screens submitted source to restrict accepted commands and imports, limiting access to functionality beyond the intended scope of introductory proof development.

These measures reduce the exposed functionality but do not establish formally verified isolation. PwnEscape (CVE-2026-99999) illustrates the dependence of VM isolation on surrounding host software: a Linux networking flaw enabled reported guest-to-host escape on the researchers’ test systems [13]. The reported virtual-network attack path is absent from our configuration, but other host-guest interfaces remain part of the trusted execution boundary.

Model Serving and Credential Handling

The default deployment serves Qwen3-Coder-30B-A3B-Instruct [14] through llama.cpp on a separate inference server. This backend reflects the current resource budget and can be replaced without changing the prover integrations.

Requests to both the default and external providers pass through the manager, which provides common handling of document context, streamed responses, and proposed source edits (Figure 2). This centralized design also requires the manager to handle the API keys supplied by BYOK users.

The security breach at OpenCode Console in October 2026 highlights the risks of handling users’ API keys [15]. To limit credential exposure, ProofStudio retains keys only temporarily in server memory, without persistent storage, and ends retained access when the associated session expires. The default session lasts 100 minutes; this design still requires users to trust the service while their keys are in use. Its privacy policy, which users must accept before submitting proof documents, also recommends disabling or rotating API keys with their provider after use [16].

Standby Allocation Under Resource Constraints

To reduce startup delays, the default configuration for a server with 64 GiB of memory shares four configurable slots between user sessions and standbys. As Figure 3 shows, with n∈{0,…,4}n \in\{0,\ldots,4\} allocated sessions, including those connecting, the target is min⁡(3,4−n)\min(3,4-n) standbys, prioritized as Isabelle, HOL Light, then HOL4. Isabelle and HOL Light initialize both the VM and prover; HOL4 initializes only the VM. Lean and Rocq start on demand because their prover initialization is comparatively inexpensive.

Desired allocation under the current resource-constrained default

Figure 3. Desired allocation under the current resource-constrained default. White named cells are standbys; shaded cells are user sessions of any prover. Rows show targets after replenishment, not instantaneous readiness or persistent slot identities. Booting and teardown also occupy capacity.

An arrival takes a matching standby, waiting if it is still booting, or starts a new VM. Only unused standbys may be reclaimed; assigned sessions are protected. Released VMs are destroyed and replacements start from a clean state. Replenishment requires memory headroom, boots one standby at a time, and waits for a default 30 s interval without new arrivals or Check/Prove jobs. Standbys serve interactive sessions only; these explicit jobs have no separate standby pools by default. The policy reduces startup work but cannot guarantee a warm instance for every arrival.

Discussion and Conclusion

Related work. Generic proof interfaces predate ProofStudio: Proof General provides a common interface to diverse proof assistants [17]. Browser interfaces also have an established history. ProofWeb supports logic teaching through Coq [18]; jsCoq runs Coq in the browser [19]; Clide connects web-based editing to Isabelle’s asynchronous document model [20]. Lean4web offers a server-backed Lean editor for small developments and demonstrations [21]. ProofStudio combines five

Data Availability

The public service, in-interface usage guidance, and ten example proof scripts for each of the five provers are available at https://proofstudio.org/. Implementation source is available to reviewers upon request. We intend to release the entire software stack as open source. However, recent security incidents, including the OpenCode Console breach [15], have heightened our concern about protecting users’ API credentials. We are therefore postponing public source release until after internal penetration testing.

native prover environments, corresponding examples, preconfigured automation, and AI assistance in one service.

Early use and proposed evaluation. A doctoral student with a mathematics background is exploring ProofStudio to learn Lean. This informal use has not produced systematic evaluation results. Classroom use is under discussion; we welcome collaboration with educators and researchers. A proposed study would ask learners to inspect an example’s specification, explain an intermediate goal, and complete a proof hole, then compare a corresponding example in another prover. Task completion, time, errors, and short explanations of the specification would inform usability and comprehension assessment. A classroom deployment should use a dedicated instance provisioned for its expected concurrency.

Limitations and future work. The current scope is single-document developments with the library coverage in Table I. Public-service capacity is limited. Future work includes multi-document projects, broader libraries, and increased capacity. Further development and evaluation would benefit from research collaboration and additional computational resources.

Conclusion. ProofStudio makes five proof assistants accessible through a common browser workflow. Users can explore the same algorithmic question across provers, inspect intermediate proof states, and review AI suggestions against native prover feedback. The service complements dedicated prover IDEs by lowering the barrier to exploration while keeping specifications and proof evidence visible.

Acknowledgment

We used ChatGPT, Claude, Codex, Claude Code, OpenCode, and Kimi-K3 to assist with drafting and revising passages throughout the manuscript, implementing ProofStudio, and preparing and producing the demonstration video. We used Lean, HOL Light, HOL4, Rocq, and Isabelle/HOL for proof development and demonstrations. The author takes responsibility for the final content. This work was supported by the research programme of the Strategy AV21 AI: Artificial Intelligence for Science and Society, with institutional support RVO:67985807.

References

  1. [1]Anthropic, “LLMs are making formal software verification feasible at a much larger scale,” Oct. 2026, published 22 October 2026. Accessed 24 October 2026. [Online]. Available: https://www.anthropic.com/research/formal-verification
  2. [2]OpenAI, “Sharing AI progress in mathematics,” Oct. 2026, published 6 October 2026. Accessed 24 October 2026. [Online]. Available: https://openai.com/index/sharing-ai-progress-in-mathematics/
  3. [3]———, “Mathematics repository: Release history,” 2026, update of 7 October 2026. Accessed 24 October 2026. [Online]. Available: https://github.com/openai/math/blob/main/history.md
  4. [4]A. Bastounis, F. Circelli, and A. C. Hansen, “Navier–Stokes lost in translation: Why Lean verification of AI formalisation does not guarantee correct natural language proofs,” 2026, arXiv:2610.08144, preprint. [Online]. Available: https://arxiv.org/abs/2610.08144
  5. [5]Lean Prover Community, “How to install Lean 4 on macOS,” https://github.com/leanprover-community/leanprover-community.github.io/blob/603abca a741aa3460039ae86351f49fbaaf4ffc1/templates/install/macos.md, 2023, historical installation instructions, revision of 7 September 2023. Accessed 23 October 2026.
  6. [6]Lean FRO, “Manual installation,” https://lean-lang.org/install/manual/, see the instructions for installing the Lean 4 VS Code extension. Accessed 23 October 2026.
  7. [7]———, “Frequently asked questions,” https://lean-lang.org/faq/, section “Is it safe to build Lean code and libraries?” Accessed 23 October 2026.
  8. [8]Y. Nagashima and D. S. Goc, “Abduction prover in Isabelle/HOL,” arXiv:2606.04877, 2026, preprint. [Online]. Available: https://arxiv.org/abs/2606.04877
  9. [9]J. Limperg and A. H. From, “Aesop: White-box best-first proof search for Lean,” in Proceedings of the 12th ACM SIGPLAN International Conference on Certified Programs and Proofs. ACM, 2023, pp. 253–266.
  10. [10]Ł. Czajka and C. Kaliszyk, “Hammer for Coq: Automation for dependent type theory,” Journal of Automated Reasoning, vol. 61, pp. 423–453, 2018.DOI
  11. [11]T. Gauthier, C. Kaliszyk, J. Urban, R. Kumar, and M. Norrish, “TacticToe: Learning to prove with tactics,” Journal of Automated Reasoning, vol. 65, pp. 257–286, 2021.DOI
  12. [12]T. Nipkow, Ed., Functional Data Structures and Algorithms: A Proof Assistant Approach, ser. ACM Books. ACM, 2025.DOI
  13. [13]PWNAI Research and P. Yibeli, “PwnEscape: Guest-to-host KVM escape (CVE-2026-99999),” https://www.pwn.ai/blog/pwnescape, 2026, accessed 21 October 2026.
  14. [14]Qwen Team, “Qwen3-Coder-30B-A3B-Instruct: Model card,” https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct, 2025, accessed 21 October 2026.
  15. [15]OpenCode, “Security breach affecting OpenCode Console users,” Announcement on X, Oct. 2026, posted 17 October 2026. [Online]. Available: https://x.com/opencode/status/2111519637049319818
  16. [16]ProofStudio, “Privacy policy & terms of use,” 2026, policy version 2026-09-18.3. Accessed 23 October 2026. [Online]. Available: https://proofstudio.org/privacy
  17. [17]D. Aspinall, “Proof General: A generic tool for proof development,” in Tools and Algorithms for the Construction and Analysis of Systems, ser. LNCS, vol. 1785. Springer, 2000, pp. 38–43.DOI
  18. [18]M. Hendriks, C. Kaliszyk, F. van Raamsdonk, and F. Wiedijk, “Teaching logic using a state-of-the-art proof assistant,” Acta Didactica Napocensia, vol. 3, no. 2, pp. 35–48, 2010.
  19. [19]E. J. Gallego Arias, B. Pin, and P. Jouvelot, “jsCoq: Towards hybrid theorem proving interfaces,” Electronic Proceedings in Theoretical Computer Science, vol. 239, pp. 15–27, 2017.DOI
  20. [20]C. Lüth and M. Ring, “A web interface for Isabelle: The next generation,” in Intelligent Computer Mathematics (CICM 2013), ser. LNCS, vol. 7961. Springer, 2013, pp. 326–329.DOI
  21. [21]Lean Prover Community, “Lean4web: Lean web editor,” https://github.com/leanprover-community/lean4web, 2026, accessed 16 October 2026.

Paper details

vlnc:W_a7b8ke5p
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

ChatGPT, Claude, Codex, Claude Code, OpenCode, Kimi K3 (bare)

View statement
Looking for related papers…
Contents