Your workflows, made agentic.Your runtime, made reliable.

The distributed runtime for agents that touch production.

Apache-2.0 · install in seconds — or join the cloud waitlist.

kortecx :: runtimebooting
Runtime booting; capacity registered.
$kx run workflow
[+0.000s]runtime.bootkortecx ready

The execution engine for agentic intelligence.

Not a database, not a model, not a chat app — the execution kernel beneath them. Describe the work; the runtime takes over.

▸ ▢▸ ▢▸ ▢
dispatch
work fans out
▢✕▢✓
recover
replay, never rerun
▸▸▸
deliver
exactly once
observe
every step traced

The model runs where the agents run.

Local GPU inference and frontier-on-demand routing live in the same substrate that schedules agents — the model thinks, emits tokens, and the agent acts, all under one runtime and one policy layer.

Crash it. Replay it. Same digest.

Every step is one content-addressed mote, appended to a journal that is the single source of truth. Kill the process mid-commit and replay: committed work is re-read, not re-run — exactly-once delivery with transparent recovery.

Read the quickstart →

Knows the actions. Re-runs on new inputs.

The runtime remembers the plan a task follows. Feed it updated inputs and it re-runs that plan dynamically — adapting to the change instead of burning tokens.

$kortecx::task memory
planbuild → test → reportknown
inputsv2 · updatedre-run
result#c204fa6d… committed
same plan · new inputs · re-run dynamically

Every kind of data, content-addressed.

Morphic captures it all in one fabric — and gives every piece an identity derived from its content. It's a projection of the journal, so it's reproducible by reference: lose it, and the runtime re-derives it.

$morphic::data fabric
content
vectors
models
recipes
tools
context
memos
content-addressed
#a6b5c679
journal-authoritative
reproducible by reference

Import an asset. Run it straight away.

Recipes, tools, models, and datasets are content-addressed bundles. Pull one by handle and invoke it — or share it: grant another identity access and they run it from their own runtime, remotely, and get byte-identical results.

$kx catalog ls
  • recipekx/recipes/map_reduceshared
  • recipekx/recipes/react_tool_loop
  • toolkx/tools/web.fetch
  • modelkx/models/qwen3-4b
  • datasetkx/datasets/fixtures
  • recipekx/recipes/retry_until_critic
kortecx :: share & invoke
your runtime
kx/recipes/map_reducegrant → id:team
teammate's runtime
$ kx invoke kx/recipes/map_reduce
▸ committed :: a6b5c679… · byte-identical

One runtime, three shapes.

Adding distribution is wiring, not a rewrite — the scheduler and executor are the same from your laptop to distributed capacity. Drop a node and watch the runtime re-lease its work exactly once.

mode:
runtime.topology::local
drop a node:
Runtime in local shape — the scheduler and executor are unchanged across shapes.

Local

One machine. Install, describe the work, full exactly-once guarantees on your laptop.

placement
single process
delivery
exactly-once
recovery
local journal replay
best for
first runs · daily work

The same binary runs in every shape. Distribution is configuration — placement is least-loaded and starvation-free, and a lost node's work is recovered without re-running anything already committed.

Policy-gated by default.

Safety isn't prompt engineering. Every action runs inside a scope the runtime enforces — in-scope work passes, out-of-scope work is refused before it touches the world, with the reason on the record.

Refused before it runs

Out-of-scope actions are refused before they ever reach a tool — not logged after the fact.

Scope only narrows

As work fans out, each step can only ever be more restricted than its parent. Never wider.

Critics gate mutations

Deterministic critics — schema, bounds, dedup — gate every world-mutating step. Non-override.

$kx --policy::requested action vs. granted scope
fs.read./workspace/*checking
model.invokesummarizequeued
net.egressapi.kortecx.comqueued
fs.write/etc/hostsqueued
db.writeno idempotency tokenqueued
shell.execrm -rf /queued

Summon agents for Engineering at Enterprise scale.

Each agent runs under a scoped warrant — only the permissions its job needs. The runtime places it, and when one crashes it re-leases the work and recovers it exactly once.

$kx summon--role engineering8 live
atlasbuild
fs.readfs.writeshell.execnet.egress
running
junotest
fs.readshell.exec
running
nyxreview
fs.readnet.egressmodel.invoke
running
vegadeploy
net.egresssecrets
running
orionmigrate
net.egresssecrets
running
echoscan
fs.readmodel.invoke
running
sableindex
fs.readnet.egresssecrets
running
kestrelsummarize
fs.readmodel.invoke
running
scoped by warrant · placed least-loaded · recovered on loss · exactly-once delivery

Built for individuals. Engineered for enterprises.

Easy enough to install and run on day one. Robust enough to run a business on. The same runtime, whether you're automating your own work or running agents across a team.

For individuals

Start in minutes

Install the runtime, describe what you want done, and step away. No infrastructure to set up, no expertise required. The runtime picks sensible defaults — you focus on outcomes, with full exactly-once guarantees on your own machine.

For enterprises

Scales with you

Reliable execution, transparent recovery, and delivery guarantees come with the runtime — not bolted on later. Capacity grows with your workload. Security, observability, and policy are designed in from day one.

$kx rundata.pipelinePer-run sandbox. Only what's needed. Released when done.
deps::pythonpytorchpandasduckdbcudanodetypescriptfaissnumpymatplotlibsqlglotnetworkxgodockerkubernetespostgressparkrayterraformrequests
[+0.012s]sandbox.create::isolated container, fresh state
[+0.041s]deps.resolve::3 needed, 0 extra
[+0.087s]deps.install::3 installed, lean image
[+0.123s]task.ready::policy-gated, observed
[+0.156s]task.complete::sandbox released in 14ms

Run your agents on a runtime that recovers itself.

Install locally and start in minutes — Apache-2.0, built in the open. Exactly-once delivery, transparent recovery, policy-gated by default.

or join the cloud waitlist for managed capacity.