Shannon Prover
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@Shannon ProverProve the 'encryption_secrecy' lemma."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Shannon Prover
From a security theorem to an EasyCrypt proof.
Shannon Prover uses language-model agents to construct machine-checked cryptographic proofs in EasyCrypt. You supply the formal description of the scheme, its security assumptions, and the theorem you want to prove. Shannon Prover automates the search for both the proof decomposition and the proofs of its individual lemmas. It can introduce intermediate games and auxiliary lemmas, try to prove them, and revise the argument as EasyCrypt exposes what remains to be shown.
Paper: ShannonProver: Towards Automating Formal Cryptographic Proofs
Website: project overview and proof browser
Contact: shannonprover@gmail.com · GitHub
What does Shannon automate?
Think of formalizing a cryptographic security proof in three phases:
Phase | Work to do | With Shannon Prover |
I — State the security problem | Define the scheme, adversary, security games, assumptions, and final theorem in EasyCrypt. | You provide these. |
II — Build the security argument | Choose intermediate games, formulate auxiliary lemmas, and connect them to the final theorem. | Shannon proposes and revises this decomposition. |
III — Prove each step | Write the detailed EasyCrypt proofs of the lemmas and the final theorem. | Shannon searches for proof scripts and checks them with EasyCrypt. |
Phases II and III need not happen in a single forward pass. For example, while trying to justify a game hop, Shannon may discover that a proposed invariant is insufficient. It can refine the auxiliary lemma, try its proof again, and then return to the main argument. This feedback between designing the argument and proving its steps is what we mean by interleaved.
One agent works on the overall argument. It can write proofs directly and ask other agents to work on selected auxiliary lemmas. Their results feed back into the overall proof, which is checked again as a complete EasyCrypt file.
flowchart LR
Input["You: scheme, assumptions,<br/>security games and theorem"] --> Plan["Shannon: propose games<br/>and auxiliary lemmas"]
Plan --> Proof["Shannon: write and check<br/>EasyCrypt proofs"]
Proof -->|"results and remaining goals"| Plan
Proof --> Final["Check the complete<br/>EasyCrypt development"]Related MCP server: math-logic-mcp
Install
Use macOS or Linux with Git, opam, Python ≥ 3.12, and uv.
Install the system build tools and native libraries before bootstrapping EasyCrypt. On Ubuntu/Debian:
sudo apt-get update
sudo apt-get install -y git curl opam build-essential autoconf automake m4 pkg-config \
libgmp-dev libpcre2-dev zlib1g-dev unzipOn macOS, install Xcode Command Line Tools (xcode-select --install if needed).
If you use Homebrew, install the native dependencies with:
brew install opam autoconf automake pkgconf gmp pcre2 zlibFor other package managers, see the vendored EasyCrypt installation guide. Shannon's bootstrap builds the OCaml dependencies but expects these system dependencies to be installed already.
1. Python environment
Install uv outside this project's .venv, then synchronize the locked Python
environment:
git clone https://github.com/SkyShannonProver/shannon-prover.git
cd shannon-prover
uv sync2. SMT solvers
EasyCrypt's smt tactic calls external solvers. Install
Z3 and CVC5 and put
their executables on your PATH before bootstrapping. CI uses Z3 4.15.4 and
CVC5 1.2.1. Check that both commands work:
z3 --version
cvc5 --versionThe next step installs Alt-Ergo 2.4.3 in Shannon's managed opam switch.
3. Repository-managed EasyCrypt
Shannon Prover is locked to EasyCrypt r2026.06. The bootstrap command creates
and verifies the repository-managed opam root and switch:
uv run python tools/bootstrap_easycrypt.py
SHANNON_OPAM_ROOT="$(uv run python -c \
'from core.easycrypt.toolchain import managed_opam_root; print(managed_opam_root())')"
SHANNON_OPAM_SWITCH="$(uv run python -c \
'from core.easycrypt.toolchain import managed_switch_name; print(managed_switch_name())')"
opam install --yes --root "$SHANNON_OPAM_ROOT" \
--switch "$SHANNON_OPAM_SWITCH" alt-ergo.2.4.3
uv run python tools/bootstrap_easycrypt.py
uv run python tools/bootstrap_easycrypt.py --verify-onlyThe solver configuration lives in .toolchains/opam-r2026.06/why3.conf.
Installing Alt-Ergo can cause opam to rebuild EasyCrypt dependencies. The second
bootstrap command records and verifies the resulting installation as well as
refreshing its solver configuration.
Worktrees of one repository share this configuration; independent clones have
their own. Bootstrap and runtime checks preserve your global Why3 configuration.
After changing solvers on PATH, rerun --configure-solvers. After package
changes in the managed opam switch, rerun the full bootstrap instead. Existing
Shannon installations need --configure-solvers once to create the local config
if their recorded toolchain is otherwise unchanged.
Python entry points select this environment and configuration automatically.
For a developer command that invokes easycrypt directly, export the verified
environment and pass the configuration explicitly:
eval "$(uv run python tools/bootstrap_easycrypt.py --print-env)"
easycrypt -why3 "$SHANNON_WHY3_CONFIG" config--verify-only and runtime availability checks read the configuration without
regenerating it.
4. Agent login
Project-level proving defaults to OpenAI Codex with gpt-6-astra (high
reasoning effort) for both the overall argument (outer) and auxiliary lemmas
(inner). Install Codex and authenticate with a stored ChatGPT OAuth login.
The commands below use this pair without additional provider flags. Claude
Code remains available when explicitly selected for either role.
See the project-level proving guide
for model settings, login requirements, and time limits.
Project-level proving
This workflow covers Phases II and III: Shannon constructs the proof decomposition and the proofs needed to reach your project's final theorem.
Put your EasyCrypt target and its local dependencies in one directory under
projects/. The file should contain the security model and final theorem,
with an unfinished proof. You do not need to supply a decomposition into
auxiliary lemmas.
projects/my-proof/
Target.ec
Dependency.ec
interleaved_project.jsonCreate interleaved_project.json with the following contents, replacing the
file and theorem names with yours:
{
"schema_version": 1,
"target_file": "projects/my-proof/Target.ec",
"final_lemma": "TargetSecurity",
"include_dirs": ["easycrypt-src/theories", "projects/my-proof"],
"artifact_root": "artifacts/my-proof-interleaved",
"max_parallel": 2,
"max_inner_minutes": 120,
"outer_timeout_seconds": 43200
}This allows at most two auxiliary-lemma attempts at once, up to two hours per attempt, and a twelve-hour overall task. These are time limits, not a total spending cap. Adjust them before running.
Commit your project files and configuration: the runner requires a clean Git working tree, including for the preliminary check. Then, from the repository root, check the setup without starting a proof-generating model:
uv run python -m workflow.interleaved \
--project projects/my-proof/interleaved_project.json \
--preflight-onlyIf the check passes, run the same command without --preflight-only:
uv run python -m workflow.interleaved \
--project projects/my-proof/interleaved_project.jsonShannon works on the target file in your checkout and prints the directory containing the run's results. The project-level proving guide explains how to read the outcome, continue partial work, and add checks that protect your original theorem and assumptions.
For a supplied cryptographic task, follow the ChaChaPoly example. It includes a security model and target theorem, without a human-supplied decomposition, and uses stricter checks that preserve the supplied problem.
What counts as a proof?
For an interleaved run, success requires a fresh EasyCrypt check of the
complete target file and no remaining admit. in that file. The run's
manifest.json records status: "proved" and
final_verification_passed: true; final_verification.json contains the
final checking result. A completed auxiliary lemma alone does not establish
the final theorem. A partial proof is reported as incomplete.
EasyCrypt checks the statements in the submitted development under its assumptions. You should also review that these are the security statement and assumptions you intended. The generic checker does not freeze the original definitions or prohibit new assumptions; projects needing those restrictions must supply additional checks. The ChaChaPoly example supplies them for its fixed task.
Single-lemma proving
This workflow covers Phase III for a lemma you have already stated.
The standalone Codex prover defaults to gpt-5.6-sol with high reasoning
effort. Project-level proving uses gpt-6-astra with high reasoning effort
for both outer and inner agents, as described above. Both workflows default
to Codex; their model settings are separate.
If you already know the decomposition, you can ask Shannon to prove
one lemma. Put its EasyCrypt file and local dependencies under projects/,
leaving the selected lemma with an unfinished proof, for example:
lemma my_lemma : true.
proof.
admit.
qed.Run Codex from the repository and enter $prove my_lemma. In the Codex
desktop app, type / and choose Prove from the skills list.
It searches projects/ first, then the bundled eval/examples/.
For example, the supplied PIR lemma can be launched with:
$prove PIR_correctClaude Code has the corresponding command:
/prove PIR_correctThese commands prove a selected lemma; they do not start the interleaved project-level workflow. The Codex command uses Codex, and the Claude command uses Claude. A generated proof is written back only after a fresh EasyCrypt verification succeeds. See the Prove instructions for the explicit-file command and other details.
Examples and research evaluation
The proof browser shows recorded single-lemma attempts and their checked steps. It does not yet show the complete sequence of game choices, auxiliary-lemma attempts, and revisions in an interleaved run.
Research evaluation is a different workflow
You do not need eval_suite to run an ordinary proof on your own project.
Controlled evaluations additionally hide existing answers and fix the task,
models, and budgets so results can be compared.
Strict live evaluation currently requires Linux and bubblewrap when using
the isolated benchmark suite; the ChaChaPoly reference has its own documented
source-access policy. See the benchmark guide and
ChaChaPoly reference guide.
Further reading
Project-level proving guide: project setup, agents, limits, and results.
ChaChaPoly example: reproduce the supplied security-proof task.
Architecture: optional implementation details for contributors.
License and citation
Shannon Prover is released under the MIT License. The vendored EasyCrypt source has its own MIT license. If you use Shannon Prover in research, please cite CITATION.cff:
@article{ma2026shannonprover,
title = {ShannonProver: Towards Automating Formal Cryptographic Proofs},
author = {Ma, Yiping and Tsai, Yu-Lin and Rathee, Mayank and Rathee,
Deevashwer and Dupressoir, Fran\c{c}ois and Strub, Pierre-Yves
and Popa, Raluca Ada},
journal = {arXiv preprint arXiv:2607.02847},
year = {2026}
}This server cannot be deployed
Maintenance
Related MCP Connectors
MCP server for building and testing AI agents with multi-model experimentation and insights.
LLM Orchestration Agent (Cryptography)
MCP server connecting AI agents to 100+ apps (Gmail, Slack, Notion, GitHub) via one-click OAuth.
Related MCP Servers
- FlicenseNot gradedqualityDmaintenanceA demonstration server that allows large language models to perform penetration testing tasks autonomously by interfacing with the Mythic C2 framework.79-
- AlicenseAqualityDmaintenanceMCP server that gives small LLMs verified symbolic-math & logic tools.62Apache 2.0
- AlicenseAqualityDmaintenanceAn MCP server that exposes the Prova reasoning verifier, enabling AI agents to verify their own reasoning and kernel-check Lean 4 proofs before outputting answers.5MIT
- AlicenseAqualityCmaintenanceA server that enables LLMs to securely manage and execute commands on remote SSH servers.643 PyPI3MIT