BootProof is designed for the moment every developer knows: you find a serious open-source repo, clone it, and then spend the afternoon guessing how to make it run.

This demo shows BootProof treating a Supabase-style stack as something that must be proved, not assumed. It infers the stack, identifies the boot path, starts the services, verifies localhost health, and writes a signed attestation.

What this proves: BootProof does not just print a happy message. It only gives the green check after observed health evidence exists.

AI coding agents can suggest commands, but they should not be trusted to declare success. This demo shows the BootProof agent loop on a GitLab-style repo: AI suggests a repair, BootProof requires approval, runs one bounded step, reruns verification, and writes a receipt showing what changed.

What this proves: AI can help move the repo forward, but BootProof refuses to fake a green check. No proof, no green check.

Attestations are JSON. That is the right shape for machines and CI, but it is the wrong shape for a human you are trying to convince. The Living Receipt is the same evidence, rendered as a single self-contained HTML file you can email, attach, or link from a README badge.

The two receipts in the demo below are not mock data. They are real captures from a real bootproof up run against two real local repositories — one that boots to a real HTTP 200, and one that installs clean and then segfaults at runtime. The signature, the boot logs, and the observed HTTP response are all real captured evidence, signed with ed25519 and embedded in the file.

Open it anywhere — no install, no network, no account — and four things happen in your browser, with zero external calls:

1. It re-verifies its own ed25519 signature. The signed message, the signature, and the public key are all embedded inline. Native WebCrypto Ed25519 is used in Chrome 137+, Firefox, and Safari; on older browsers (~20% of users) the receipt automatically falls back to a pure-JS @noble/curves verifier bundled inline. The receipt never fails to verify silently.
2. It replays the actual boot — the inferred plan, the real commands that ran, the real captured log timeline, the real observed HTTP response (or the real classified failure) — streaming in an embedded terminal.
3. If a single byte of the signed message is altered, the verdict collapses with the signature. The boot stamp flips to VERDICT UNVERIFIED — SIGNATURE INVALID, because no claim inside a tampered receipt can be trusted. That is the whole point of the product: a green check that survives tampering would defeat the entire premise.
4. The trust ladder is documented in the artifact itself. This receipt is signed at the local_developer_signed level — it proves integrity-since-signing, not that the signer's machine was honest. The documented upgrade path is local_developer_signed → ci_oidc_signed → neutral_runner_signed → transparency_logged. The same file format and the same wow experience survive every step up the ladder; only the signer's identity climbs.

Try it now — download the Living Receipt and open it in your browser:

curl -sL https://github.com/bootproof/bootproof/raw/main/assets/living-receipt.html -o proof.bootproof.html
open proof.bootproof.html   # macOS
# or: xdg-open proof.bootproof.html   # Linux
# or: double-click the file in Finder/Explorer

Then click Tamper with signature to watch the verdict collapse. The self-verification only works when the file is opened locally (file://) — GitHub shows the raw source inline, which is why you need to download it first.

_{(right-click → Save Link As… → save as .html → double-click to open)}

Reproduce the capture. The engine that produced these real receipts ships in this repo — both the standalone MVP (scripts/bootproof_up.mjs) and the integrated TypeScript CLI:

# Using the TypeScript CLI (the published npm package) — emits the receipt natively
npx bootproof up <any-repo> --provider local --unsafe-local --install --receipt

# The receipt is written to .bootproof/living-receipt.html

# Or use the standalone MVP engine (for development/testing)
node scripts/bootproof_up.mjs fixtures/real-booting-app --label "real-booting-app"

# Regenerate the Living Receipt HTML from MVP captures
node scripts/build_living_receipt.mjs \
  scripts/records/real-booting-app.json \
  scripts/records/real-slop-app.json \
  --out assets/living-receipt.html

# Run the smoke test (verifies both native and fallback paths)
node scripts/verify_living_receipt.mjs

--receipt is now a native flag. npx bootproof up <repo> --receipt runs the real TypeScript CLI, boots the repo for real, observes health for real, signs the attestation with ed25519, and emits the self-verifying Living Receipt HTML — all in one command. The receipt contains the real captured evidence from that run, not mock data.

What this proves: the receipt is not a description of a product — it is the product, and it runs the moment you open it. It works offline, survives being forwarded, and cannot be forged without breaking the signature. A platform-issued green check dies the moment you leave the platform. A Living Receipt travels.

The PLG loop is wired in. Every receipt now carries:

• A first-time visitor banner ("You're holding a Living Receipt — forward it, it verifies itself on the next machine")
• A "Copy markdown badge" button that generates a [![bootproof](...)](./proof.bootproof.html) snippet for this specific receipt's verdict
• A "Download this file" button — the file IS the product, one tap to share it
• A page-level CTA: npx bootproof up <any-repo-url>

Every receipt is a self-proving advertisement. Every badge in the wild is a link. Every click is a developer asking "what is this?" — and the answer is the product. See assets/bootproof-badge-template.md for copy-paste badge snippets, and docs/LAUNCH_PLAYBOOK.md for the "First 1,000 Receipts" distribution sequence.

The honest run button for GitHub repos. Proof, not vibes.

BootProof answers one question:

Did this repository actually boot?

Not “did a command run?”
Not “did Docker say containers are up?”
Not “did an AI agent say it worked?”
Not “did the README look plausible?”

BootProof inspects a repo, builds an evidence-based run plan, executes only what it can justify, observes real health, and writes a signed attestation for success or failure.

No proof, no green check.

Every developer knows this loop:

git clone some/repo
npm install
npm run dev

Then reality appears.

Wrong Node version. Wrong pnpm version. Missing Java. Missing Clojure. Docker is running but the service is not healthy. Postgres exists but the role does not. Redis is missing. A migration fails. The app starts but nothing responds. A container is “up” but the product is unusable. An AI agent confidently says “done” because a process started.

That is not proof.

BootProof exists because repo onboarding should not depend on hope, terminal archaeology, or fake green checks.

Modern repositories are no longer simple.

A repo might contain:

• multiple workspaces
• Docker Compose services
• frontend and backend apps
• hidden runtime requirements
• package-manager version constraints
• generated assets
• database migrations
• health endpoints
• undocumented local assumptions

A README can be useful, but it is not proof.

A terminal command can be useful, but it is not proof.

A model response can be useful, but it is not proof.

BootProof turns repo booting into an evidence trail.

BootProof separates activity from evidence.

A failed run is still useful if it tells the truth.

✗ NOT VERIFIED — package_manager_version_mismatch

What happened:
The repository requires pnpm 10.24.0, but this environment has pnpm 9.15.4.

Why BootProof refused:
The dependency install cannot be trusted with the wrong package manager version.

Safe next step:
Run corepack enable && corepack prepare pnpm@10.24.0 --activate, then rerun BootProof.

Evidence:
.bootproof/attestation.json

Predictable failure is a feature.

Run BootProof against a local repo:

cd /path/to/repository
npx bootproof up .

BootProof will inspect the repo and either prove it booted or explain why it refused.

For CI or agent workflows:

npx bootproof up . --ci --json

For explicit local execution:

npx bootproof up . --provider local --unsafe-local

Run dependency installation only when you intend to:

npx bootproof up . --provider local --unsafe-local --install

Explain or verify an attestation:

npx bootproof explain .bootproof/attestation.json
npx bootproof verify .bootproof/attestation.json

BootProof can inspect public HTTPS repositories from GitHub, GitLab, Bitbucket, and Codeberg.

npx bootproof up https://github.com/dubinc/dub

Remote repositories are untrusted code, so BootProof inspects first and refuses execution until you explicitly opt in.

Remote source: https://github.com/dubinc/dub.git
Clone retained at: .bootproof/remotes/github.com/dubinc/dub-*/repo

Inference
  application: yes
  package manager: pnpm
  selected command: pnpm dev

✗ NOT VERIFIED — remote_code_execution_blocked

Why BootProof refused:
Remote repositories are untrusted code and require explicit consent.

To run remote code locally, you must say so explicitly:

npx bootproof up https://github.com/dubinc/dub --provider local --unsafe-local --install

BootProof never silently executes remote code.

✓ install: dependencies installed
✓ start-app: app process started and was supervised
✓ health: observed HTTP 200 at http://localhost:3333

✓ BOOTED — HTTP 200 at http://localhost:3333

Evidence:
.bootproof/attestation.json

A repository is only marked BOOTED when BootProof observes health evidence.

A process start is not enough.
A successful install is not enough.
A Docker container is not enough.
A command exiting is not enough.

Humans get a readable diagnosis:

NOT VERIFIED — workspace_ambiguous

BootProof detected a root command that starts multiple workspaces in parallel.
Choose a specific application with --workspace <dir>; one responding workspace
is not proof that the whole repository booted.

Example:

npx bootproof up . --workspace apps/studio

BootProof is designed to make failures legible.

It should tell you whether the problem is:

• package-manager mismatch
• skipped install
• missing runtime
• ambiguous workspace
• unsupported orchestration
• allocated port
• failed service
• failed app start
• unhealthy endpoint
• health timeout

--json emits exactly one bootproof/result/v1 object:

{
  "schema": "bootproof/result/v1",
  "booted": false,
  "healthVerified": false,
  "failureClass": "dependency_install_skipped",
  "attestationPath": ".bootproof/attestation.json",
  "inference": {},
  "plan": {},
  "observed": []
}

--ci disables colour and interactive prompts.

Exit codes are deterministic:

That makes BootProof useful for CI, agent workflows, and repo-quality gates.

BootProof has been tested against real repositories, including small apps, monorepos, large platforms, and multi-service stacks.

The point is not to turn every repo green. The point is to produce the correct verdict.

Failure is not hidden or relabelled as support.

Evidence stays evidence.

See docs/REAL_REPO_EVIDENCE.md.

A fresh BootProof run against supabase/supabase detected:

stack: make-driven, node-frontend, docker-compose
repo compose: docker/docker-compose.yml
workspaces: apps/studio, apps/www, apps/docs, packages/*
selected command: make dev

BootProof refused:

✗ NOT VERIFIED — workspace_ambiguous

The root command starts multiple workspaces in parallel.
One responding workspace would not prove that the whole repository booted.

That refusal is correct.

Manual follow-up through Supabase’s official Docker route proved the platform path:

cd docker
cp .env.example .env
docker compose up -d

Core services such as Kong, Studio, DB, Auth, REST and Pooler reported healthy/running, and localhost:8000 returned Kong/API responses.

The lesson:

BootProof should not fake a monorepo-wide success just because one endpoint responds.

Airbyte correctly exceeded BootProof’s direct orchestration boundary.

BootProof refused instead of pretending a normal Gradle, Make, or Compose command was enough. The documented local path required abctl, Kind and Helm. A human followed that runbook and booted the application.

BootProof could then verify the external health endpoint without claiming it started Airbyte.

bootproof verify-url http://localhost:8001/api/v1/health

External verification means:

This endpoint responded.
BootProof did not orchestrate the startup.

That distinction matters.

npx bootproof up .

npx bootproof up . --workspace apps/studio

npx bootproof up . --ci --json

npx bootproof up . --provider local --unsafe-local

npx bootproof up . --provider local --unsafe-local --install

npx bootproof verify-url http://localhost:8001/api/v1/health

npx bootproof up . --external-health http://localhost:8001/api/v1/health

npx bootproof explain .bootproof/attestation.json

npx bootproof verify .bootproof/attestation.json

npx bootproof diff --base main --head HEAD
npx bootproof diff --base main --head HEAD --json

npx bootproof fix .

OPENAI_API_KEY=... npx bootproof fix . --ai

or:

ANTHROPIC_API_KEY=... BOOTPROOF_AI_PROVIDER=anthropic npx bootproof fix . --ai

bootproof up remains zero-AI.

BootProof is built for a world where humans and AI agents both touch repositories.

The intended loop is:

Diagnose
→ Classify
→ Plan
→ Risk-classify
→ Approve
→ Execute one step
→ Verify
→ Receipt
→ Repeat

AI can suggest.
Humans can approve.
BootProof proves.

The complete autonomous loop is not implemented. Today BootProof exposes four honest modes.

bootproof up .

BootProof infers a supported local run path, executes it within the selected safety boundary, observes health, and writes an attestation.

Unsupported or ambiguous orchestration is refused.

bootproof verify-url http://localhost:8001/api/v1/health

BootProof observes a service started outside BootProof. Successful evidence is classified as externally verified and never claims BootProof started the app.

bootproof plan-agent .

BootProof writes a deterministic, risk-classified plan and a redacted local receipt chain. It does not execute candidate actions, and planning never counts as success.

bootproof fix .

BootProof maps exact known failures to deterministic repair actions. Mutating commands and patches require explicit approval. Verification decides whether the failure progressed or the application booted.

See:

• docs/AGENT_IN_THE_LOOP.md
• docs/AGENT_RUN_RECEIPTS.md
• docs/DETERMINISTIC_REPAIR_SAFETY_MODEL.md

bootproof fix reads the latest signature-valid classified failure and maps exact known failures to deterministic actions.

bootproof fix .

Host and service commands show the exact command, scope and risk. They run only when the user gives explicit approval. JSON and CI modes never approve commands.

Repair receipts distinguish:

• declined
• failed
• progressed
• verified

Machine mode:

bootproof fix . --json

It emits one bootproof/repair-result/v1 object and exits 0 only when a verified receipt exists.

fix never applies file patches directly. To explicitly apply a signature-valid file repair to a local working tree:

bootproof apply-repair .

Application checks the receipt signature, allowed file scope, signed content hashes, and exact current preimages before writing.

See docs/REPAIR_RECEIPT.md.

AI suggestions are optional and are only available after no deterministic repair is known.

OPENAI_API_KEY=... bootproof fix . --ai

or:

ANTHROPIC_API_KEY=... BOOTPROOF_AI_PROVIDER=anthropic bootproof fix . --ai

BootProof asks before contacting the provider, sends only redacted structured failure evidence, validates the strict bootproof/ai-repair-suggestion/v1 response through the shared safety model, and asks again before any command or patch is tested.

AI suggestions are recorded as ai_suggested.

They never enter the deterministic registry automatically.

bootproof diff --base main --head feature-branch
bootproof diff --base main --head HEAD --json

diff reads committed Git objects and performs static analysis only.

It does not:

• check out either ref
• execute repository code
• install dependencies
• read protected .env contents
• upload data

It reports supported drift in:

• dependency manifests and lockfiles
• Compose services and ports
• environment variable names
• start commands
• package managers
• runtime markers
• detectable health routes

A diff can require fresh proof, but it never claims the head revision boots. Run bootproof up against the intended revision to establish that with observed health evidence.

BootProof is constrained on purpose.

It will not:

• mark a repo BOOTED without observed health
• execute remote code without explicit consent
• fall back from Docker to host execution silently
• render skipped steps as success
• invent secrets
• write protected .env files
• silently patch project code
• guess a workspace when the repo is ambiguous
• claim generated scaffolding exists unless it was written
• upload telemetry or hidden evidence

It will:

• sign successful attestations
• sign failed attestations
• preserve local evidence
• classify known failures
• refuse unsupported paths clearly

See docs/HONESTY_CONTRACT.md.

BootProof's execution model is honest by default, not isolated by default. There is no general-purpose container sandbox in the current release. Here is exactly what happens when you run bootproof up:

• Default (--provider docker): BootProof will only execute inside Docker for source-built Compose applications (repos with a docker-compose.yml where the app service is built from source). For every other repo — plain Node, Python, Rust, Go — the Docker provider refuses to run with orchestration_not_supported rather than silently falling back to the host. This is intentional: the default is fail-closed, not silent host execution.
• --provider local --unsafe-local: runs install and start commands directly on your host machine using spawn(command, { shell: true }). There is no container, no network restriction, no read-only filesystem. The --unsafe-local flag is the explicit consent gate — you are acknowledging that you have reviewed the inferred commands and accept that the repo's code (including postinstall scripts, prestart hooks, and anything the start command does) will run on your machine with your privileges.
• Remote repos (bootproof up https://github.com/...): BootProof clones for inspection but refuses to execute without --provider local --unsafe-local. Inspection is safe; execution requires consent.

Before running bootproof up --provider local --unsafe-local on a repo you didn't write:

1. Run bootproof up <repo> --dry-run first to see the inferred commands without executing them.
2. Read the plan. The install command and start command will run on your host.
3. Only then add --unsafe-local --install to actually execute.

This is the current truth. General-purpose Docker isolation for non-Compose repos is on the roadmap but is not in this release. If that's a blocker for your use case, do not use BootProof on untrusted repos yet.

BootProof treats repair actions as executable risk.

The repair safety model blocks or escalates dangerous commands before they run.

Blocked examples include:

• sudo
• shell interpreters
• pipe-to-shell downloads such as curl | sh
• inline arbitrary execution such as node -e, python -c, ruby -e
• recursive world-writable chmod
• raw disk writes
• destructive database drops
• protected .env writes
• secret exfiltration patterns

High-risk actions require explicit approval and can never be downgraded by AI-provided risk labels.

See docs/DETERMINISTIC_REPAIR_SAFETY_MODEL.md.

BootProof currently provides:

• Node package-manager and start-command inference
• monorepo candidate ranking
• Docker service dependency detection
• repository Compose detection
• conservative Go main-package execution
• Rails bin/rails entrypoint detection
• explicit Make run-target execution
• Python/Flask and Go/Node hybrid detection
• localhost health-candidate discovery from repo evidence and app logs
• classified failures
• signed Ed25519 attestations
• strict JSON and fail-closed CI output
• redacted registry-entry export
• deterministic sandboxed repairs for registered failure classes
• explicit repair application with signature, scope and stale-preimage checks
• static infrastructure diff

Detection is broader than orchestration. BootProof may detect a stack and still refuse to run it if the proof boundary is not safe or clear.

Supported execution paths are deliberately narrow.

Each path still requires observed health.

A successful docker compose up -d, process spawn, or command exit is not a green result by itself.

Common failure classes include:

• not_an_application
• workspace_ambiguous
• dependency_install_skipped
• package_manager_version_mismatch
• python_flask_setup_required
• service_port_allocated
• postgres_auth_env_missing
• health_http_error
• health_check_timeout
• remote_code_execution_blocked
• unknown_failure

Unknown failures remain unknown, with evidence preserved for the next detector.

See docs/FAILURE_TAXONOMY.md.

Depending on the command and observed plan, BootProof may write:

.bootproof/attestation.json
.bootproof/registry-entry.json
.bootproof/registry/<timestamp>-<hash>.json
.bootproof/runtime/
docker-compose.bootproof.yml
.env.bootproof.example

Registry artifacts are written only by explicit export commands.

Protected application env files remain untouched.

Local attestations (the default) contain:

{
  "trust": {
    "level": "local_developer_signed",
    "signer": "local_ed25519",
    "oidc": null
  }
}

Local attestations are useful evidence. CI/OIDC attestations are stronger supply-chain proof. BootProof does not pretend local laptop proof is enterprise CI proof.

Inside GitHub Actions with permissions: id-token: write, pass --ci-oidc to fetch the runner's OIDC token and embed its claims in the attestation:

bootproof up . --provider local --unsafe-local --ci-oidc

The attestation then carries ci_oidc_signed trust with the OIDC claims:

{
  "trust": {
    "level": "ci_oidc_signed",
    "signer": "local_ed25519",
    "oidc": {
      "iss": "https://token.actions.githubusercontent.com",
      "sub": "repo:bootproof/bootproof:ref:refs/heads/main",
      "repository": "bootproof/bootproof",
      "run_id": "123456",
      "workflow": "CI",
      "job_workflow_ref": "bootproof/bootproof/.github/workflows/ci.yml@refs/heads/main"
    }
  }
}

The ed25519 signature still provides integrity; the OIDC evidence provides CI provenance. A verifier can independently validate both the signature and the OIDC claims. The neutral_runner_signed and transparency_logged levels remain on the roadmap.

The local signing key can be rotated without invalidating existing attestations (each carries its public key inline and verifies independently):

bootproof rotate-keys                    # generate new key, back up old
bootproof rotate-keys --repo . --resign  # also re-sign the latest attestation

Old keys are archived to ~/.bootproof/archived-keys/ so existing attestations remain verifiable.

BootProof does not upload attestations.

A project can deliberately export a redacted local registry entry or federated public-candidate receipt and review it before committing it.

bootproof registry export .
bootproof attest export .
bootproof registry export . --federated

Public crawler, private cloud upload, and OIDC-backed trust are future integrations, not deployed services in this repository.

See:

• docs/CI_ACTION.md
• docs/REGISTRY.md

This repository contains the local trust layer:

• local diagnosis
• local planning
• local receipts
• local approvals
• optional BYOK AI suggestions
• deterministic repair safety
• no telemetry
• no automatic upload

The OSS engine works offline and does not require BootProof Cloud.

BootProof Cloud belongs in a separate private repository.

Its boundary includes future hosted capabilities such as:

• hosted AI
• shared registry
• team approval workflows
• GitHub App
• SSO/RBAC
• policy
• fleet dashboards
• audit retention

These are product boundaries, not claims that those services are implemented in this public repository.

No Cloud/SaaS code is included here.

The npm package contains the compiled CLI, license, README and docs.

dist/ is required at runtime, generated by npm run build during prepack, and intentionally not committed.

Run:

npm run pack:check

This packs BootProof, installs the tarball in an isolated temporary directory, and exercises the installed CLI.

See docs/RELEASE_CHECKLIST.md.

For contributors working from source:

git clone https://github.com/bootproof/bootproof.git
cd bootproof
npm ci
npm run build
npm test
npm link

Then from another repository:

bootproof up .

Generated files such as dist/, node_modules/, and .DS_Store are ignored and not committed.

BootProof is not:

• a deployment platform
• a general CI replacement
• a magic environment fixer
• an AI coding agent
• a guarantee that every repo can be run automatically
• a cloud product hidden inside an OSS repo
• a sandbox or container runtime — --provider local runs code on your host; --provider docker only isolates source-built Compose apps

BootProof is the honest run button for repos.

It runs what it can, refuses what it cannot prove, signs both success and failure, and gives humans and machines the same evidence.

BootProof is early alpha.

Near-term work includes:

• explicit full-platform Compose mode
• stronger multi-service health modelling
• broader deterministic remediation coverage
• more Python, Go, Ruby and Make entrypoints
• CI/OIDC-backed signing
• proof-linked badges
• verified public evidence index

Unsupported paths should fail clearly, not magically.

Apache-2.0