Documentation

Everything you need to run the kernel: getting started, guides for each part of the system, and a full reference for executors, guards, stores, and validation.

Start here

Introduction — what Praxec is and the problem it solves.
Installation — install the praxec kernel and the px control-plane TUI.
Quick start — your first unbreakable rule, on the wire, in about a minute.
How to run it — the ways to run the kernel and what each one enforces.

Guides

Agent configuration (agents.yaml) — Set up per-affinity model bindings once, share them across workflows. Closed enums, sparse overrides, eager auth preflight.
Capabilities and orchestrators — Praxec's two-tier composition model: typed capability sub-workflows you compose into lifecycle orchestrators. Write a capability once. Use it everywhere.
Checking your workflows — Prove a config loads, behaves, and holds its declared properties with praxec check, fuzz, and test -- no model, no network, CI-ready.
Composing a small architecture from scratch — Walk through building a minimal code-review architecture: two skills, one workflow, two agents, ready to ship.
Connections — Your agent can call create_pull_request directly today. Connections plus a guard are how you put that behind a tests-passed gate — wiring in MCP servers, CLI commands, and REST APIs as backends.
Deterministic chaining — Tag transitions the LLM doesn't need to decide and the runtime chains through them in a single round trip.
Discovery & search — How your model finds capabilities without a bloated tool list -- search, describe, and follow links.
File locks & coordination — A global repo lock so no two agents write the same file at once. Contention is never an error — a blocked workflow durably suspends and auto-resumes the moment its files free.
Going to production — What to change before your gateway handles real traffic -- storage, audit, validation, and operational concerns.
Guardrails: moves your agent can't make — Stop your coding agent shipping on a red suite or approving its own PR. Earned-move gates, expression and evidence checks, and human-only moves -- declared in YAML, enforced before any command runs.
Hot reload — Reload your gateway config with SIGHUP -- no restart, no dropped workflows.
How scripts work — A script turns a check like cargo test into a step the agent must pass through, not a suggestion it can skip — and pins the body so it can't drift. Inline body or external uri+hash, invocable from any workflow. Here's the operational rundown.
How skills work — A skill is a rule the agent reads at the moment it acts — your review checklist, your 'never force-push main' discipline — not prompt text it skims and forgets. The operational rundown: where skills live, how they load, and how the gateway hands them over on demand.
In-runtime LLM execution — Host the model call inside the runtime with kind: llm. The transitions available at the current state become the model's entire tool list — under enforced caps, with every turn audited.
Indirection & caching — Two fixed tools means the model discovers capabilities instead of reading them all up front. Here's what that costs, what's cached, and the honest floor.
Multi-repo loading — Load any number of resource repos via a top-level repos: block. Every loaded definitionId is namespace-prefixed; collisions fail at load; explicit overrides shadow vendored content.
Phase guidance — Give the model goal and reasoning instructions at each workflow state so it knows what to focus on and how to think about choices.
Self-authoring with praxec-meta — Five meta-orchestrators that use the framework to author content for the framework. The library inspects your reachable tooling, proposes 2–3 shapes, and emits validated YAML before the PR opens.
Skills, workflows, and cognitive architectures — Your agent's full config, split into reusable layers: what it can do, the order it does it in, who does each step, what it's allowed to connect to.
What's inside a workflow library — Skills, scripts, capabilities, orchestrators, connections, shared context — the pieces a reusable workflow library packages, how they compose, and why bundling them in one Git repo is the point.
Wire praxec into your editor — Wire praxec as an MCP server into Zed, Cursor, Claude Desktop, Claude Code, and VS Code (via Continue/Cline). Copy-paste config for each editor.
Workflows — State machines that decide which moves your agent can make and in what order -- from flat tool calls to multi-step pipelines.

Documentation

Start here

Guides

Reference

Advanced