hc-stark
Pre-release MIT resource-bounded Plonky3 0.6.1 prover with SSD-backed execution and deterministic resume for Linux x86-64.
Documentation
TinyZKP
TinyZKP is an MIT-licensed, resource-bounded proving backend for Plonky3. Its
production objective is simple: let proving teams complete larger STARK traces
under an explicit RAM ceiling, using deterministic SSD scratch where global
transforms require it, while retaining the official Plonky3 proof format and
unmodified verifier.
> Release authority: the MIT Community source is public; a production
> engine binary is supported only when its automated release evidence passes.
> TinyZKP does not operate hosted proving, hosted verification, customer
> accounts, usage metering, or MCP commerce. Guard artifacts and checkout are
> independently fail-closed by the evidence-derived
> `release/guard-launch-state-v2.json`,
> generated from
> `release/guard-launch-evidence-v2.json`
> and verified against protected signing and launch trust. Consult those
> generated contracts for current availability; do not infer it from README
> prose.
Product boundary
TinyZKP does not introduce a new production transcript, verifier, proof format,
or security profile. The production path is pinned to Plonky3 `0.6.1` and the
upstream Goldilocks/Poseidon2 `p3-uni-stark` example configuration, frozen as
`tinyzkp-p3-goldilocks-v1`.
The differentiating layer is prover-side infrastructure:
- resource policies for RAM, scratch, threads, and checkpoint retention;
- checksummed, owner-only matrix stores and block-readable matrix access;
- a deterministic scratch-backed Plonky3 DFT adapter;
- official Plonky3 proof generation and verification for Fibonacci and
Goldilocks Poseidon2 reference AIRs;
- twelve frozen Guard-facing contracts generated from the shared
`tinyzkp-contracts` Rust authority, plus proof-engine contracts;
- Linux cgroup-v2 measurement from fresh process creation through verification;
- release provenance, compatibility locking, and fail-closed publication gates.
The resource-bounded pipeline is implemented locally: trace and quotient
generation, four-step transforms, MMCS, openings, and FRI use durable bounded
stores, and typed checkpoints preserve the official challenger state for
byte-identical resume. Plonky3 `0.6.1` still hands its generic DFT trait an
owned `RowMajorMatrix`; TinyZKP's bounded orchestration therefore uses the
separate block-matrix entry point. The implementation remains pre-production
until the automated verifier, determinism, fixed-host resource, recovery,
fuzz, provenance, SBOM, signature, CLI, and OCI checks pass for the exact
release and the owner-controlled Guard launch gates publish that artifact.
Independent reviews and design-partner results remain visible advisories.
Standalone TinyZKP protocols, legacy receipts, hosted services, recursion, zkML, zkVM, IPA,
Spartan, KZG, and rollup prototypes are research only. Legacy CLI access
requires the explicit `legacy-research` feature and is not compiled into the
public engine container or default CI path.
Repository map
| Path | Purpose |
|---|---|
| `crates/tinyzkp-contracts` | Frozen public JSON contracts, reason vocabulary, schemas, and resource arithmetic shared by the engine and Guard |
| `crates/hc-stream` | Resource policy, block matrices, matrix stores, preflight, and checkpoint contracts |
| `crates/hc-plonky3` | Pinned Plonky3 configuration, workloads, DFT adapter, official prover/verifier, and artifact contracts |
| `crates/hc-cli` | Production Plonky3 CLI and benchmark worker |
| `scripts/benchmark` | cgroup-v2 benchmark orchestration and report tooling |
| `release` | Dependency compatibility and mandatory release-gate evidence |
| `site` | Static Guard product, compatibility, evidence, documentation, and legal-status site |
| `docs/recovery` | Architecture, delivery, release, and operating documentation |
Server, MCP, billing, SDK, and hosted-beta sources remain temporarily on this
branch pending the obligation inventory, verified final exports, replacement
live smoke, and authorized decommission. Their workflows are disabled and
their binaries are excluded from active release and deployment payloads. The
`archive/hosted-beta-2026-07-17` tag preserves the historical snapshot; it is
not evidence that the remaining source or external infrastructure has already
been removed.
Build and test
Rust `1.95.0` is pinned because Plonky3 `0.6.1` uses APIs unavailable on older
toolchains. Plonky3 and artifact-serialization dependencies are exact-pinned in the
workspace and verified against `release/plonky3-compatibility-v1.json`.
cargo test --locked -p hc-stream -p hc-plonky3 -p hc-cli
cargo clippy --locked -p hc-stream -p hc-plonky3 -p hc-cli \
--all-targets -- -D warnings
python3 scripts/ci/guard_launch_gate.py --checkThe hosted API/MCP/billing launch audit is retired and cannot authorize Guard.
The signed engine candidate, Guard candidate, static site, and OCI identities
must instead pass the Guard evidence and joint promotion controls.
CLI
Generate the twelve frozen Guard-facing API schemas, the standalone
compatibility-profile schema, and the proof-engine schemas from their Rust
sources of truth:
cargo run --locked -p hc-cli -- schema --output-dir /tmp/tinyzkp-schemasThe installed production-facing executable is `tinyzkp-engine`; `hc-cli` is the
Cargo package and development binary name. Run the compatibility doctor with a
populated `JobManifestV1` before proving:
tinyzkp-engine doctor --job job.jsonThe declarative AIR proof path uses an exact caller-owned checkpoint directory:
tinyzkp-engine plonky3 prove-air \
--air \
--trace-manifest \
--chunks-dir \
--public-inputs \
--policy \
--checkpoint-dir \
--output
tinyzkp-engine plonky3 inspect-checkpoint \
--checkpoint \
--air \
--trace-manifest \
--chunks-dir \
--public-inputs \
--policy
tinyzkp-engine plonky3 resume-air \
--air \
--trace-manifest \
--chunks-dir \
--public-inputs \
--checkpoint \
--output
tinyzkp-engine plonky3 verify-air --bundle`inspect-checkpoint` performs the same identity and durable-artifact checks
without changing job state. `resume-air` then validates every checkpoint
identity and durable artifact,
reconstructs the uploaded declarative workload, restores the official
challenger state, and continues from the last completed phase. Exact-release
crash/resume tests require the resulting proof bytes to match an uninterrupted
run exactly.
Generic `prove` and `verify` commands return migration guidance. Historical
reproduction is available only in an offline research build:
cargo run -p hc-cli --features legacy-research -- legacy-research --helpThe installed `tinyzkp-engine release` command emits JSON for cross-checking
the engine binary, OCI image, compatibility profile, and benchmark provenance
before publication. In a source-development checkout, its internal equivalent
is `cargo run --locked -p hc-cli -- release`.
Benchmark integrity
Release measurements must run on Linux under cgroup v2, with baseline and
candidate each launched in a fresh process. The report includes hardware, OS,
storage, release SHA, dependency profile, exact command, CPU seconds,
whole-process peak memory, scratch high-water mark, block I/O, proof size,
verification time, and verifier result.
cargo build --release -p hc-cli
sudo python3 scripts/benchmark/run_plonky3_cgroup.py \
--manifest examples/plonky3/fibonacci-small.json \
--report /tmp/fibonacci-bounded.jsonmacOS measurements and component-only tests are useful for development but are
not acceptable release evidence. TinyZKP never infers full-prover memory,
throughput, cost, or capacity from a component benchmark.
Release targets remain blocked until the owner-signed automated suite proves:
- 1M rows: at least 4× lower peak RAM, at most 3× baseline wall time, and no
more than 10% above the configured cap;
- 16,777,216 rows: at most 2 GiB whole-process peak memory, successful official
verification, and scratch usage within 10% of preflight;
- deterministic crash recovery, parser/resource fuzzing, signed artifacts,
SBOM, checksums, provenance, CLI/OCI smoke, and release identity agreement.
Independent reproduction, outside review, and external workloads are
transparent `not_completed` advisories and do not authorize checkout.
Self-hosted behavior
The public product is a local binary and customer-operated OCI image. It has no
job API, account database, queue, worker fleet, proof storage, usage meter, or
runtime TinyZKP dependency. Customer witnesses, scratch data, and proofs remain
on customer-controlled storage.
The separate commercial Guard supervisor may activate a signed release through
the merchant-of-record. After activation, doctor, run, resume, policy, and
verify operations are offline. Cancellation prevents activation of future
releases but does not disable an already activated release.
Commercial model
Proof-critical software remains MIT. TinyZKP intends to sell one
customer-operated product:
- Community: free MIT engine, verifier, schemas, doctor, reference workloads,
and public evidence.
- TinyZKP Guard: $499/month or $4,990/year for one legal organization's
internal use, with automatic mode selection, recovery supervision, CI policy,
signed qualification, and access to new releases.
There is no free Guard trial, usage metering, Enterprise tier, Fleet/OEM plan,
hosted compute, custom AIR work, SLA, onboarding call, or included engineering.
The free engine and doctor are the evaluation path. Public checkout remains
disabled until the technical, legal, merchant, external-workload, unaided
installation, and first-customer gates are evidenced.
The supported compatibility profile remains fixed. A different profile may be
considered for a quarterly qualification window only through the local,
aggregate-only demand gate documented in
`docs/validation/PROFILE_EXPANSION_DEMAND_GATE.md`;
meeting that demand threshold does not make the candidate supported.
The machine-readable commercial source is `site/pricing.json`.
Commercial launch state is generated by
`scripts/ci/guard_launch_gate.py` from
reviewed V2 evidence; the derived state file is not a manually editable launch
approval.
See `BUSINESS_GUIDE.md` for operating controls and
`docs/recovery/implementation-status.md`
for the current gap ledger.
Operators must follow
`docs/runbooks/release_provenance.md`,
`docs/validation/FOUNDING_VALIDATION_PROTOCOL.md`,
and
`docs/runbooks/guard_commerce_setup.md`.
Optional counsel intake is consolidated in
`docs/governance/GUARD_COUNSEL_PACKET.md`;
that packet is advisory and is not required launch authority. LN Holdings'
owner approval of exact seller facts and document digests is the machine legal
gate.
Security and disclosure
Do not commit secrets, witness data, customer inputs, payment-provider credentials,
private keys, or production environment files. Scratch artifacts must be
owner-only and are untrusted on reopen; manifests, chunks, release identity,
dependency lock, workload, input, and policy must all be validated before
resume.
Report security issues through the private channel linked from
tinyzkp.com/security. Performance and security
claims require reproducible evidence; pre-release source is not a production
certification.
License
This repository and its public engine code are MIT licensed. TinyZKP Guard is a
separate, commercially licensed supervisor. Guard must invoke the public engine
through the published file and CLI contracts; it may not fork proof semantics
or place proof-critical behavior behind the commercial license.
Frequently asked questions
What is hc-stark?
hc-stark is Pre-release MIT resource-bounded Plonky3 0.6.1 prover with SSD-backed execution and deterministic resume for Linux x86-64.
How do I install hc-stark?
Open the GitHub repository and follow its README. Most MCP servers are added to your client's MCP config, then called by your agent.
Is hc-stark open source?
Yes — it is hosted on GitHub at https://github.com/logannye/hc-stark.
Related MCP tools
A super light-weight embedded code search engine CLI (AST based) that just works - improves speed and efficiency for coding agent 🌟 Star if you like it!
Cognee is the open-source AI memory platform for agents. Give your AI agents persistent long-term memory across sessions with a self-hosted knowledge graph engine.
Automate browser based workflows with AI
Hindsight: Agent Memory That Learns
A privacy-first app that strips AI watermarks from content you own.
Agent framework and applications built upon Qwen>=3.0, featuring Function Calling, MCP, Code Interpreter, RAG, Chrome extension, etc.
Run your own MCP server? See who uses it and what to fix.
Measure it with TrackMCP