// CoreLink — content-addressed cache & execution fabric

Work is a pure function of its inputs. So do it once.

CoreLink is a multi-tenant, content-addressed cache and execution fabric — live on Cloudflare's network across five production regions — where identical work is never repeated and compute is a disposable cursor over a shared store. The bill follows the structure: flat concurrency instead of metered minutes, zero egress instead of transfer fees.

Software re-does finished work constantly: the same dependency compiled on a million machines, the same node_modules filling a million disks, CI re-testing code nobody touched. Give every piece of work a content address, do it once, share it. Watch it run below — the logs are real.

Live · self-serve · no card to start
corelink — the journey, in six acts
$ clw snapshot --name demo-journey .   # tree → content-addressed snapshot
$ clw hydrate /tmp/fresh --name demo-journey
$ clw -v run --input src -- cargo test # memoized: unchanged inputs never re-run
content-addressed on zero-egress R2 — identical work is stored once and shared across the org clients Bazel · Turborepo sccache · agents CoreLink Cache CAS + Action Cache isolation model-checked in TLA+ consumers Runners · Workspaces ephemeral compute
the moat · global cache

The internet's packages, downloaded once — for everyone.

npm, PyPI, Homebrew and OCI — every public wheel, bottle and layer is fetched from upstream once, ever, then served from a shared, content-addressed mirror across every tenant. Your private work stays walled off in TLA+-proven isolation; the public commons is shared.

npmPyPIHomebrewOCI
fetched from upstream once, ever
CoreLink _public mirrorcontent-addressed · token-gated · zero-egress R2
served to every tenant
acmeglobexinitechyou+ next
The flywheel a single-tenant cache can't have: every new customer leaves the mirror more complete for the next one. More teams → a fuller cache → faster and cheaper for everyone on it. Private deps stay isolated behind cryptographic tenant walls; public deps compound into a commons that only gets better with scale.
Drops into what you already run Bazel REAPI v2Turboreposccachenpm · pip · brewDocker / OCInative CAS & AC
the platform

One product. Three surfaces over one shared store.

One bet, applied three ways: if work is a pure function of its inputs, then the cache, the CI runner, and your working tree are all just addresses over the same content-addressed store.

live · the store

Your organization's shared cache — every artifact, stored once.

Inside your tenant, every artifact is addressed by what it is (BLAKE3), not where it lives — so your whole team holds one copy of everything, shared across every laptop, CI job and region. A teammate's build becomes your instant hit; CI's compile lands warm on your laptop. Private, isolated, and yours.

Content-addressed store

CAS + Action Cache keyed by BLAKE3/SHA-256. Identical inputs resolve to identical outputs — computed once, anywhere, reused everywhere. Your tools already speak it: Bazel REAPI v2, Turborepo, sccache.

Isolation, proven in TLA+

Cross-tenant byte non-leakage is a formal spec that runs in CI on every change. Each blob path is HMAC-derived over tenant identity under a key the edge can't read — another tenant's path can't even be guessed.

Action Cache, not just blobs

Memoize whole build actions, not only their outputs — an unchanged action never re-executes anywhere in your org. Native Bazel REAPI v2 and Turborepo remote caching, with a per-tenant $-ceiling that fails closed. (Public packages get their own shared mirror — below.)

Zero-egress, multi-region

Stored on Cloudflare R2 across five regions — pulling from the cache is free, exactly where S3-backed caches bleed on transfer fees. Residency is enforced per tenant, not assumed.

corelink-cas · tenant acme
$ bazel build //... --remote_cache=…humangr.com
  1,284 artifacts · 1,190 hits (92.7%)
$ pip install --index-url …/pip/acme numpy
  numpy ← _public mirror (upstream once, ever)
  ✓ tenant path = HMAC(tenant) · un-guessable
  ✓ audit  ed25519 chain · verified offline
TLA+ isolation specs in CI · BYOK across 4 KMS · per-tenant $-ceiling, fail-closed · Ed25519 audit log, verified offline · zero-egress R2
live · the compute

CI that stops repeating itself — and stops billing by the minute.

Runners boot cache-warm and replay memoized steps instead of re-running them. Fresh, fail-closed microVMs; cryptographically signed outcomes; and a bill that goes flat instead of counting minutes.

Cache-warm boot

Inputs are already local when the job starts, and memoized steps replay instead of re-running. The repeats — the compile nobody changed, the test nobody touched — cost nothing.

Fail-closed microVM

Each job gets a fresh microVM, destroyed afterward. Images are digest-pinned and refused unless verified; boxes carry per-lease scoped tokens, never master credentials.

Signed attestation

Every result returns as an Ed25519-signed attestation binding the full outcome — inputs, image, result — verifiable offline, with no trust in the runner required.

Flat, not metered

Flat per parallel runner, unlimited minutes — modeled well under hosted per-minute CI at typical use, before memoization zeroes the repeats entirely.

corelink-runner · job #4821
# job picked up — boots cache-warm
▸ inputs      local · 0 bytes downloaded
▸ microVM     fresh · fail-closed · digest-pinned ✓
# memoized steps replay; unchanged never re-runs
✓ cargo build replayed from memo · 1.2s
✓ cargo test  ok. 214 passed; 0 failed
▸ teardown    microVM destroyed · token expired
✓ attestation ed25519:9f3c1a… outcome bound
cap-safety proven live · advisory-lock admission control · recursive adversarial audit → zero P0/P1 · per-lease scoped tokens
live · the working tree

Your working tree becomes an address.

clw snapshots a directory into the store and hydrates it on any machine in seconds — only the chunks that machine is missing travel. Heavy checkouts stop duplicating themselves across every laptop and runner.

Snapshot → one ref

clw snapshots a whole tree into a single content-addressed ref. The working tree stops being a place on a disk and becomes an object you can move by name.

Hydrate anywhere

Only the chunks a machine is missing travel — a fresh box pulls from zero-egress R2, a warm one from local L1. Seconds, not a full re-clone, on every machine.

Memoized runs

clw run memoizes by inputs — unchanged inputs never re-run, they replay the recorded result. Compute becomes a disposable cursor you point at work.

Signed & shipped

Shipped to Homebrew, five-target releases minisign-signed, per-crate SBOM, a local L1 cache. The journey in the hero ran these exact commands.

~/box · clw
$ clw snapshot --name feature-x .
  root d82d924b…685043be · 3 chunks up
$ clw hydrate ~/box --name feature-x
  bytes_from_cache 205 · downloaded 0
$ clw run --input src -- cargo test
  [clw] cache hit · replayed in 1.64s
  test result: ok. 1 passed; 0 failed
Homebrew · minisign-signed five-target releases · per-crate SBOM · local L1 cache · chunk-level dedup
5
production regions
TLA+
isolation, proven in CI
Ed25519
signed audit chain
$0
egress on cache pulls

Stop paying to re-upload the same bytes.

Spin up a tenant, point your build cache at CoreLink, and watch the cache-hit rate climb — with the tenant isolation and audit trail your security team will actually sign off on.

Get started free →