Use Hyperstruck Core from the AI tools you already use. This repository brings together lightweight ways to adopt the platform without shipping a heavy SDK: starting points you can copy into your project or publish internally.
Putting domain expertise, knowledge, and memory to work is what turns AI into real business outcomes. Core delivers advanced reasoning — structured plans, milestones, and steps that steer work toward clear goals. The approach is grounded in evidence, designed to be safe, auditable, and secure, and paired with cognitive learning so you get quality over quantity: less noise, more trust, and time back for your team.
Practical learning accumulates what works: from your organization, from operators, and from knowledge you provide. That learning feeds the reasoning layer — including resolving conflicting facts, applying lessons to future plans, and improving over time. You can manage learnings yourself (for example from an IDE) or rely on automated flows inside the platform.
| Path | What it is for |
|---|---|
| claude_skills/hyper-reasoning/ | Hosted reasoning from Claude Code or Cursor: send rich context, receive structured plans and analysis, handle review checkpoints when policies require it. |
| claude_skills/hyper-learning/ | Learnings API from your agent: search before hard tasks, store insights after, reinforce what helped — so the platform remembers and ranks knowledge appropriately. |
| claude_skills/hyper-plans/ | Plans API from your agent: search similar plans and review candidate learnings surfaced with those results. |
| hyperstruck-py/ | The hand-written hyperstruck Python package: a thin HTTP client plus an optional LangGraph middleware (the langgraph extra). The ergonomic learning surface, install with pip install hyperstruck[langgraph] and your agent learns run over run. |
| sdk/ | Generated low-level API clients (Python and TypeScript) produced from the platform's OpenAPI schema. Raw, fully-typed access to every endpoint; prefer hyperstruck-py for the ergonomic learning surface, reach for these when you need an endpoint the package does not wrap. |
Together, these skills mirror how Core is meant to be used: reason when problems are deep or cross-cutting, learn continuously so the next session starts smarter.
Skill names use the hyper- prefix so they do not clash with third-party skills in a shared skills directory. See claude_skills/README.md for a short directory overview.
There are no extra scripts or package dependencies. Instructions tell your AI assistant how to call the Hyperstruck HTTP API using built-in capabilities (for example curl or WebFetch).
The utility value returned in a learning's standing is Core's derived,
recency-weighted application-outcome score. It changes as later runs show that a
learning helped or misled an agent; it is not a confidence field or the
persistence source of truth. The optional utility accepted when manually
storing a learning is only its starting prior.
Core schema v5 is the required persistence and corpus-bundle contract. Normal API and SDK users do not construct that internal payload: use the documented learning request and response models instead. Platform-operated corpus export, import, and clone validate the complete selected corpus or bundle before producing output or writing data. A legacy or malformed learning therefore fails the operation instead of yielding a partial bundle or partial import. Existing operator-managed corpora must complete the global v4-to-v5 migration before transfer.
Copy or symlink each skill folder into your environment’s skill directory:
# Claude Code
cp -r claude_skills/hyper-reasoning .claude/skills/
cp -r claude_skills/hyper-learning .claude/skills/
cp -r claude_skills/hyper-plans .claude/skills/
# Cursor (optional)
cp -r claude_skills/hyper-reasoning .cursor/skills/
cp -r claude_skills/hyper-learning .cursor/skills/
cp -r claude_skills/hyper-plans .cursor/skills/Browsing inside the Hyperstruck Core Platform monorepo? These paths live under
public_integrations/(for examplepublic_integrations/claude_skills/hyper-reasoning).
Each skill includes:
SKILL.md— what the assistant reads on each invocation (including frontmatter where supported).reference.md— detailed request and response shapes when something more than the summary is needed.
You need a Hyperstruck API key and (for learnings) usually an agent id that scopes stored knowledge.
API key (first match wins):
- A key you paste in chat (never echoed back by the skill).
- Environment variable
HYPER_API_KEY. - A
.envfile in the project root, or a path pointed to byPUBLIC_INTEGRATIONS_ENV_FILE.
Send it on every request as:
Authorization: Bearer <your key>
API base URL — default https://api.hyperstruck.com. Override via conversation, HYPER_BASE_URL, or .env.
Agent id (for hyper-learning):
- You specify it explicitly, or
HYPER_AGENT_ID/.env, or- The skill lists available agents and asks you to choose.
Examples (exact slash syntax depends on your host):
/hyper-reasoning Map out a migration plan with milestones and risks
/hyper-learning search retry backoff
/hyper-plans search onboarding workflow
/hyper-learning store
/hyper-learning reinforce <learning-id>
hyper-reasoning is marked high effort in its frontmatter: it selects an appropriate Hyperstruck profile, assembles context from your session, submits work to Core, polls until completion, and walks through human-in-the-loop steps when a run is suspended. The first real API call is GET /agents?limit=50, which validates credentials in the same round trip as loading profiles — no extra validation request.
hyper-learning pre-approves the same HTTP tools so searches and writes do not trigger a permission prompt on every call.
Both skills can inject current environment hints at load time (for example whether HYPER_API_KEY is set) where the host supports inline shell in the skill body.
- Product and platform: hyperstruck.com
- Skill index: claude_skills/README.md
If you are extending or publishing this material from Hyperstruck’s private engineering repository, see the internal maintainer note in the platform repo: docs/public-integrations-maintainers.md.