HeepX
Build once. Learn everywhere.
Reusable AI capabilities for developers and AI agents.
$ pip install heepxWhat is HeepX
HeepX builds capability repositories: a written protocol plus a deterministic evaluation suite, meant to make an AI system better at exactly one thing. Not prompts. Not a framework. Not a product.
heepx, the Python client documented here, is an independent, unofficial way to load one — download it, verify it against a content fingerprint, and read it as a typed value. It is not itself part of the capability ecosystem.
Why HeepX exists
The problem
An AI system works something out — that a claim needs checking, that a migration step cannot be undone — and then the session ends. The next session starts from nothing. The insight isn't wrong or superseded; it simply evaporates.
The honest status
One controlled benchmark exists: a placebo-controlled delta of +0.208 at p=0.125 on one small model — not statistically significant — against a prompt-length cost significant at p=0.008. Whether loading a capability measurably helps is promising and unproven, not settled.
Capabilities
Browse all →Nine, deliberately — the ecosystem grew to nine quickly and then stopped on purpose. Each owns one boundary; none overlap.
- codingCDP-1
Bounding what a code change can affect and what may be claimed about it — enumerating the change's reach, and scoping the correctness claim to the reach that was examined and the evidence that covered it.
- coordinationCRP-1
Knowing which facts must be single across agents and whether they actually are — classifying a shared fact by how a disagreement about it resolves, and refusing irreversible action on one that has not been confirmed shared.
- memoryMDP-1
Deciding what survives across a gap you did not observe — storing only what is cheaper to recheck than to re-derive, and never claiming a verification that did not happen.
- planningPDP-1
Turning a goal into an ordered set of actions whose failure states are known — classifying each step by the recovery it admits, and ordering by what a step costs to undo rather than only by what it needs.
- reasoningRDP-1
Making the structure of reasoning inspectable — naming what kind of commitment each step is, what observation would defeat it, and what dies with it when it fails.
- securitySDP-1
Deciding what an action is allowed to do and on whose authority — classifying every input by who could have shaped it, and refusing any grant whose chain terminates in something the adversary could have produced.
- synthesisSYN-1
Combining several sources into one output without manufacturing claims — tracing every claim to a source that made it, and disposing each conflict by source quality rather than by blending it away.
- verifyVDP-1
Checking a claim against retrievable evidence — returning the smallest verdict the evidence supports, from evidence that could have come out otherwise.
- warrantEDP-1
Deciding, per claim, whether to assert, cite, attribute, verify, or abstain — by naming the claim's warrant before emitting it.
Code
All examples →Python
import heepx
warrant = heepx.learn("warrant")
print(warrant.name, warrant.version)CLI
$ hx learn warrant
warrant 1.2.1
fingerprint a1b2c3d4e5f6a7b8Offline
heepx.learn("warrant")
heepx.learn("warrant", offline=True) # no network callMulti-agent
planning = heepx.learn("planning")
verify = heepx.learn("verify")
plan = run_agent(planning, task=goal)
checked = run_agent(verify, task=plan)Architecture
Every capability owns one boundary — a specific moment where an AI system commits to something and can get it wrong. Nine repositories, none overlapping, each independently versioned and independently falsifiable.
| Repository | Boundary | The question at it |
|---|---|---|
| warrant | assertion | why is this claim true? |
| reasoning | thought | what am I building on, and what would collapse it? |
| verify | the check | does this hold against evidence? |
| planning | execution | in what order do I act, and what if a step fails? |
| memory | the session | what survives a gap I did not observe? |
| coding | the change | what can this change affect, and how far does my claim hold? |
| security | trust | what is this allowed to do, and on whose authority? |
| coordination | the agent | what do we hold in common, and has it been confirmed? |
| synthesis | the merge | what may appear in a combined output? |
Using capabilities with heepx-sdk (unofficial)
The Python client on this site is a separate, independent tool — not part of the diagram above. It resolves a name against the registry, downloads and caches the archive once, verifies its content fingerprint, and returns a typed value.
Read how the client works →Documentation
- Quick Start
Install HeepX and learn your first capability in under five minutes.
- Installation
Install the heepx package and the hx CLI.
- Python SDK
The six-function public API — overview and where the full reference lives.
- CLI
The hx command line tool — overview and where the full reference lives.
- Registry
How capability discovery works, and how to browse what's published.
- Capability Standard
The heepx.capability/v1 (HXS-1) descriptor format.
- Architecture
How the SDK's pieces — registry, downloader, cache, loader — fit together.
- Developer Guide
Contributing to the SDK: setup, checks, project layout.
- Examples
Runnable examples — overview and where the full set lives.
- API Reference
Every public SDK function and every hx command, in one place.
Roadmap
Full roadmap →Foundation
Evidence and rules
Independence, federation, durability
Community
- GitHub
- GitHub Discussions
- Discord · coming soon
- X · coming soon
- YouTube · coming soon
- LinkedIn · coming soon