01 / one skill
spend runx.ai/x/spend
expertise as a URL: the contract, package, inputs, and execution profile. nothing to rebuild by hand.
runx · the governed runtime for agent skills
plug the world into your agent
runx is the governed runtime for agent skills. It connects tools and accounts to AI agents, limits each run to the authority a user grants, and records the result in a signed receipt.
signed receipts open source · MIT built in Rust
runs inside Claude Code Cursor Codex GPT Gemini Aider
what it is
skills package expertise as links, compose under , and return proof
runx gives each skill an execution path: call the next skill, request narrow authority, and bring back receipts your team can inspect, replay, and reuse.
skill
hand your agent a skill URL from the catalog, or run a skill directly from the CLI.
authority
grant the narrow capability needed for this run: one repo, one provider, one spend cap, one approval gate.
receipt
get back the evidence with the work: inputs, scopes, approvals, outputs, artifacts, and replay context.
anatomy of a run
one goal fans out into a chain of skills
> runx skill business-ops -i renewal=failed
OKloaded skillrunx/business-ops
OKbuilt graphaccount · finance · comms · approval
OKauthority scopedcrm:read · billing:read · drafts:write
··approval gatebefore send / discount / escalation
＋ RECEIPT rx_biz…42f1 · signed
lanes, grants, gates, drafts, replay map sealed
one signal becomes coordinated work
start with one business signal. runx turns it into a governed lane of skills: account context, finance checks, draft comms, approval gates, and receipts.
because the graph is replayable, the work becomes something you can debug, rerun, improve, and feed back into future operator behavior.
the catalog
skills your agent can actually use
the catalog is the capability surface your agent can call from. connector-backed skills use only the account data and provider permissions an operator grants. inspect a skill, install the CLI, publish your own, or build against the TypeScript contracts.
connected provider example
how runx uses Google data
runx connects to Google Analytics and Google Search Console only after a user explicitly authorizes a Google account. It reads only the properties, sites, metrics, and search-performance data selected for the requested operation, then uses that data to produce analytics reports, search evidence, and SEO recommendations. Submitting a sitemap requires separate approval.
runx does not sell Google user data, use it for advertising, or use it to train a general-purpose or shared AI model. Users can revoke their connection at any time. Read the full Google data policy.
why it changes the work
three moves that make skills compound
runx turns skill knowledge into something agents can share, chain, and replay without losing the boundary.
01 / distribute
send a capability, not a briefing packet
a skill URL carries the instructions, package, inputs, and execution profile. another agent can use it without rebuilding the context by hand.
from: copied promptsto: portable skills
02 / compose
let skills fan out
a business-ops skill can call docs, release, issue, payment, notification, or publishing skills. runx records the graph and narrows authority at each edge.
from: hidden glueto: governed chains
03 / replay
bring receipts back from the chain
each consequential lane signs into the receipt chain with scopes, inputs, outputs, gates, decisions, and replay context for future runs.
from: one-shot transcriptsto: reusable workflows
how it compounds
the catalog becomes your agent's own skills
a skill is a URL. runx fans them out under authority into chains you
one-shot. then runx export installs the whole
governed catalog as native skills in the agent you already use.
02 / fan out under authority
one call fans out into a complex skill-chain graph, governed at every edge. you one-shot what used to be hand-wired.
03 / native in your agent the ultimate compound
spend sourcey deep-research issue-triage business-ops + the catalog
✦ governed by runx · every run sealed
every governed skill becomes a native skill in the agent you already use, runx export claude for example.
for teams that need proof
every run becomes reusable operating memory
signed receipts
the run record
each run records the skill, inputs, scopes, gates, outputs, lineage, and closure reason so the workflow can be replayed later.
learning loops
training examples with provenance
successful runs can become reusable examples for agents and LLMs because the context, authority, decisions, and outcomes stay attached.
operator metrics
visualize how work actually flows
project the receipt stream into cycle time, approval drag, retries, fanout depth, authority use, and skill quality without inventing a parallel event model.
self-hosted
your keys, your boundary
run it in your environment. keep credentials, policies, and private context inside your boundary.
open source
build with us
read the runtime, inspect the spec, fork the repo, or publish your own skills.
run with it
give your agent the runx skill, or install the CLI and ship your first receipted run.
https://runx.ai/SKILL.md
$brew install runxhq/tap/runx
$npm install -g @runxhq/cli
$curl -fsSL runx.ai/install | sh
the governed runtime for agent skills.
crazy skills. sane runtime.
v0.8.2 · MIT
