A semantic substrate for AI-collaborative software. Model your business. Ship your product. Pass your audit.
Dazzle is a declarative framework for building SaaS applications — and a research project exploring what software architecture looks like when AI collaborators are treated as first-class readers and writers of the codebase. You describe your business in structured .dsl files: entities, roles, rules, workflows, events. The runtime executes that description directly. There is no code generation step, no scaffold to maintain, and no second source of truth.
cd examples/simple_task && dazzle serve
# UI: http://localhost:3000
# API: http://localhost:8000/docsIf you are evaluating Dazzle rather than just trying it, start with the skeptical evaluator's guide and the security claims inventory. They show the DSL -> IR -> runtime path, RBAC matrix, runtime verifier, and compliance evidence with copy-pasteable commands and explicit limits.
Traditional software architecture is optimized for humans. Files, modules, conventions, and abstractions are arranged to fit a programmer's working memory.
That optimization is no longer the only one that matters. An increasing share of software work — exploration, modification, review, refactor, test — happens with an AI collaborator in the loop. Codebases that scatter meaning across controllers, models, migrations, and templates burn context window on plumbing the model already understands. Codebases whose intent is encoded in a typed, inspectable graph let the model spend its attention on the actual problem.
Dazzle is a bet that semantic compression — putting your application's meaning in one inspectable, machine-readable form — produces software that is easier to evolve, easier to audit, and easier to collaborate on with both humans and machines. The DSL is not a shortcut to a generated codebase. It is the codebase.
The reason the DSL is the codebase, not a generator for it, is downstream of a sharper claim: Dazzle is a prior-correction substrate for LLM-driven software development. Training corpora are dominated by popular-but-aging idioms (Rails ActiveRecord, React class components, jQuery-shaped vanilla JS, exception-as-control-flow, polymorphic associations, manual SQL string-building, untyped denormalisation). An LLM trained on that corpus has those shapes as its prior; running it as an agent at scale propagates the corpus mean into the codebase unless something else pulls against it.
The framework's job in the agent-driven era is to be that something. Dazzle implements three stacked layers that each catch what the others miss:
-
Grammar restriction. The DSL closes off bad idioms by construction. There is no
polymorphic_ref:keyword (ADR-0027). Field-level authorization isn't expressible (ADR-0025). Scope rules compile to a formal predicate algebra validated against the FK graph (ADR-0009). Regex in the parser is a smell (ADR-0024) and the allowlist sits at zero. Each closed-off shape is one degree of freedom the corpus prior can no longer exercise. -
Inference-time bias correction. Agent instruction files (
.claude/CLAUDE.md), the ADR index (each line a "decision that prevents a wrong proposal"), and the counter-prior catalogue atdocs/counter-priors/are versioned engineering artefacts that name specific corpus pathologies and route the agent toward the right shape. The catalogue is queryable via the MCP server (knowledge counter_prior query=...) and is auto-surfaced at the bootstrap step when an agent's spec text contains matching triggers. -
Post-hoc filtering. Drift gates, the conformance engine, the fitness investigator, RBAC matrix verification, and the broader test sweep catch whatever slips past the first two layers. Each gate pins a specific corpus pattern:
test_no_bare_except_pass.py,test_no_regex_in_parser.py,test_shell_strict_mode.py, the API surface drift baselines.
The catalogue is itself a deliverable. Each entry — exceptions-as-control-flow, polymorphic associations, hand-rolled soft-delete, raw SQL string-building, shell without strict mode, and the rest — is a small permanent inoculation against a recurring drift. As LLM-emitted code re-enters training corpora at scale, this counter-biasing grows in value over time: a framework that does this work today is investing in an asset that compounds across model generations.
The broader framing (corpus pathologies, compounding problem, generalised principle) is captured in the catalogue itself. The shortest version: defer decisions to runtime only when the dynamism is essential to the domain, not merely convenient for the author. Make every other decision as statically as possible, and encode the static decision in the substrate so the LLM cannot accidentally undo it.
Eleven stated positions, defended in the ADRs:
- The DSL is the source of truth. API specs, tests, compliance evidence, and runtime behaviour are all derived from the same IR.
- No code generation. The runtime executes the IR directly. No regeneration drift, no generated files to maintain.
- Anti-Turing by design. The DSL has no arbitrary computation. Everything is statically inspectable, lintable, and verifiable.
- PostgreSQL only. One capable relational database plus disciplined semantics beats distributed-systems sprawl for the workloads Dazzle targets.
- Server-rendered HTML + HTMX. No SPA framework, no build toolchain, no client-state fragmentation.
- Fragments as the only escape hatch. When the DSL can't express it, you reach for a fragment — a constrained, named, semantically-tagged piece of custom rendering. Not arbitrary frontend.
- Append-oriented history. Events, decisions, and grants are logged. Auditors don't need to spelunk; the trail is part of the substrate.
- Provable RBAC. Scope rules compile to a formal predicate algebra and are statically validated against the FK graph.
- No hidden singletons. Dependencies are explicit (
RuntimeServices,ServerState) — readable by both humans and agents. - No backwards-compat shims. Pre-1.0, clean breaks beat layered workarounds. Callers are updated in the same commit.
- Bump on every fix. Every push gets a unique semantic version — deployment traceability over release ceremony.
If you disagree with one of these, you'll probably disagree with the rest. That's the point of stating them up front.
Most frameworks ask you to express your business logic across scattered files — controllers, models, migrations, middleware, templates. When requirements change, you update all of them and hope they stay in sync.
Dazzle inverts this. You write what your business is — entities, roles, permissions, workflows, state machines — and the runtime executes it directly. Change the DSL, refresh the browser. The DSL is the single source of truth for your application, your API spec, your test suite, and your compliance documentation.
If you're building SaaS — especially in regulated industries — you will face auditors. They will ask: who can access what? how are changes controlled? where is sensitive data classified?
Most teams answer these questions retroactively, combing through code to produce evidence. Dazzle derives the answers from the DSL itself:
- Access control is declared in the DSL and provably enforced. Every permission is statically verifiable.
- State machines model approval workflows, transitions, and four-eyes authorization.
- Compliance evidence is extracted automatically. Run
dazzle compliance compile --framework soc2and get a structured audit report showing which controls your DSL satisfies. - Grant-based RBAC supports delegated, time-bounded access with approval workflows — the kind of access governance auditors want to see.
Dazzle currently supports ISO 27001 and SOC 2 Trust Services Criteria out of the box, with automatic evidence mapping from your DSL declarations to specific framework controls.
module my_app
app todo "Todo Application"
entity Task "Task":
id: uuid pk
title: str(200) required
completed: bool=false
created_at: datetime auto_add
surface task_list "Tasks":
uses entity Task
mode: list
section main:
field title "Title"
field completed "Done"Save this as app.dsl, run dazzle serve, and you have:
- A PostgreSQL-backed database with correct types and constraints
- CRUD API endpoints with pagination, filtering, and sorting
- A rendered list UI with sortable columns, search, and a create form
- OpenAPI documentation at
/docs - A health endpoint with deployment integrity verification
That's a todo app. The same language scales to 39-entity accountancy platforms with double-entry ledgers, multi-step onboarding wizards, and role-based dashboards. You add complexity only where your business needs it.
| Capability | What it does | Why it matters |
|---|---|---|
| Entities | Data models with types, constraints, relationships | Your domain model, declared once |
| Surfaces | List, detail, create, review views | UI and API from the same declaration |
| Workspaces | Role-based dashboards with filtered regions | Each persona sees what they need |
| State Machines | Lifecycle transitions with guards and approval | Business processes enforced, not just documented |
| Access Control | Cedar-style permit/forbid rules, scope predicates | Provable RBAC — auditors can verify mechanically |
| Grant Schemas | Delegated, time-bounded access with approval | Four-eyes authorization, SOC 2-ready |
| Processes | Multi-step workflows with saga patterns | Durable business operations |
| Atomic Transactions | Multi-entity writes in one scope-guarded transaction (atomic) |
No partial writes; every touched entity scope-checked, fail-closed (ADR-0029) |
| Experiences | Onboarding wizards, checkout flows | Guided multi-step user journeys |
| Ledgers | TigerBeetle-backed double-entry accounting | Financial-grade transaction integrity |
| Graphs | Entity relationships with CTE traversal and algorithms | Network analysis, shortest paths, community detection |
| HLESS Events | Intent/Fact/Observation/Derivation event semantics | Replay correctness, audit lineage, no "events as a vague bucket" |
| Fragments | Constrained custom rendering inside generated surfaces | Differentiated UX without losing semantic integrity |
| Islands | Self-contained interactive JS components mounted into server-rendered pages | Charts, editors, drag-and-drop without adopting an SPA framework |
| Integrations | Declarative API bindings with triggers and mappings | Connect to Stripe, HMRC, Xero, and more |
| LLM Jobs | Classification, extraction, generation tasks | AI capabilities without prompt engineering sprawl |
| Services | Custom business logic declared in DSL, implemented in typed Python/TS stubs | A bounded escape hatch for domain logic that keeps the declarative boundary |
| Compliance | Maps DSL constructs to ISO 27001 and SOC 2 controls | Control-coverage evidence, gaps flagged |
For the full DSL reference, see docs/reference/index.md.
A few Dazzle keywords don't map one-to-one onto industry terms. If you're skim-reading the DSL for the first time:
| Dazzle term | What other communities call this |
|---|---|
| surface | view, page, screen — a UI/API endpoint with one entity and one mode |
| workspace | dashboard, role home, console |
| experience | wizard, flow, multi-step form |
| rhythm | recurring cadence, scheduled review, periodic ritual |
| archetype | persona pattern, role family |
| hless | event-stream semantics (HLESS = High-Level Event Semantics Specification — why this name) |
| fragment | escape-hatch component, custom partial |
Dazzle treats compliance as a first-class concern, not an afterthought.
Every DSL construct that relates to security — access rules, data classification, state machine transitions, process workflows — is automatically mapped to compliance framework controls. Run:
dazzle compliance compile --framework iso27001 # ISO 27001 audit
dazzle compliance compile --framework soc2 # SOC 2 TSC audit
dazzle compliance gaps --framework soc2 # Show unmet controlsThe output is a structured AuditSpec showing which controls are evidenced (your DSL satisfies them), which are gaps (your DSL should cover them but doesn't), and which are excluded (physical security, HR — outside DSL scope).
Scope rules compile to a formal predicate algebra, statically validated against the FK graph at dazzle validate time. The verification framework has three layers:
| Layer | What it proves |
|---|---|
| Static Matrix | Every (role, entity, operation) combination is computed from the DSL |
| Dynamic Verification | The running app is probed as every role to confirm runtime matches the matrix |
| Decision Audit Trail | Every access decision is logged with the matched rule and outcome |
dazzle rbac matrix # Generate the access matrix (no server needed)
dazzle rbac verify # Verify runtime matches the matrix (CI gate)
dazzle rbac report # Compliance report for auditorsSee RBAC Verification and Compliance for details.
For apps that need it, Dazzle ships native, per-organization enterprise identity — OIDC, SAML 2.0, and SCIM provisioning — so a customer's IdP (Okta, Microsoft Entra ID, Google, Ping) drives sign-in and user lifecycle. Connections are framework-owned runtime data (not DSL), fenced to one org, and gated by DNS-verified domain ownership; secret material is AES-256-GCM encrypted at rest. SAML covers IdP-metadata import, SP-signed AuthnRequests, encrypted assertions, and bidirectional Single Logout; SCIM covers user/group provisioning with group→role mapping.
It is off by default — a greenfield app sees none of it until you opt in (dazzle capability enable auth.enterprise.oidc), so the simple path stays simple. Identity is modelled as global Identity + Organization + fenced Membership + Session, so one person can belong to many orgs with tenant isolation enforced at the data layer, not bolted on.
See Enterprise SSO & Provisioning and multi-tenant hosting.
Stating this directly because it matters:
- Real-time collaborative editing. No CRDT layer, no client-state model.
- Graphics-heavy or canvas-based interfaces. Server-rendered HTML is not the right substrate.
- Local-first or offline-first applications. Authority lives on the server and in PostgreSQL.
- General-purpose programming. The DSL is deliberately not Turing-complete. If you need arbitrary computation, you need a different tool — or you write a fragment.
- Replacing your existing codebase wholesale. Dazzle is most useful for new applications where governance, workflow, and audit are first-class concerns from day one.
The framework is strongest for enterprise SaaS, workflow systems, operational tooling, and governance-heavy applications. That's the bet.
# Install
brew install manwithacat/tap/dazzle # macOS/Linux (auto-registers MCP server)
# or: pip install dazzle-dsl
# Run the example
cd examples/simple_task
dazzle serve
# Open http://localhost:3000 for the UI
# Open http://localhost:8000/docs for the APISupported runtime: Python 3.12 – 3.14 · PostgreSQL · macOS / Linux. The floor is 3.12; CI runs the full suite on 3.12, 3.13, and 3.14 (all hard-required) on every change, and 3.14 is the primary deploy target (Heroku's default; faster on the parse path via the uv tail-call interpreter — see docs/python-3.14-primary-target.md). Development and deploys use uv (uv sync; Heroku via the native uv buildpack).
DSL Files → Parser + Linker → AppSpec (IR) → Runtime (live app)
→ OpenAPI / AsyncAPI specs
→ Test generation
→ Compliance evidence
→ Fidelity scoring
The DSL is parsed into a typed intermediate representation (AppSpec IR). The runtime executes the IR directly — no code generation step. Every artifact (API specs, tests, compliance reports, demo data) is computed from the same IR.
This architecture is deliberately anti-Turing: the DSL has no arbitrary computation, which means Dazzle can statically validate, lint, measure fidelity, and reason about your application. What you declare is what runs.
The frontend uses server-rendered HTML with HTMX — zero build toolchain, stable technology, and full visibility into what the runtime produces. For UX that the generated surfaces can't express, fragments provide a constrained escape hatch: named, semantically-tagged custom rendering that remains connected to the entity and surface graph.
For the full architecture, see docs/architecture/overview.md. For the event-semantics rationale, see docs/architecture/hless-deep-dive.md.
Dazzle ships as both a runtime and an AI development environment. When used with Claude Code (via MCP), you get access to a growing set of tools that span the full lifecycle. The exact tool count, operations, and parameters drift with development — see the MCP Tool Inventory for the live list, regenerated from the registry every build. As of the latest doc regen: 34 tools, 156 operations. Broad lifecycle coverage:
| Stage | What the tools do |
|---|---|
| Spec to DSL | Turn a natural-language idea into validated DSL — entity discovery, lifecycle identification, persona extraction |
| Test and Verify | Generate stories, design tests, execute at three tiers (API, browser, LLM-guided), seed demo data |
| Analyze and Audit | Quality pipeline, agent-powered gap discovery, visual composition analysis, RBAC policy verification |
| Site and Brand | Manage public site structure, copy, theme, and design tokens from spec files |
| Stakeholder Ops | Launch readiness scores, investor pitch generation, user/session management |
The agent framework uses an observe-decide-act-record loop to autonomously explore running applications, discover gaps, and propose DSL fixes. Discovery modes include persona-based exploration, CRUD completeness analysis, workflow coherence checks, and headless DSL/KG analysis.
For the live tool-by-tool inventory (operations, parameters, descriptions), see the MCP Tool Inventory — generated from the registry every doc build. For the architectural model, see Architecture: MCP Server. For how the autonomous slash-command harness drives day-to-day development on the framework itself, see Autonomous Harness.
# Homebrew: MCP server auto-registered during installation
brew install manwithacat/tap/dazzle
# PyPI: Register manually
pip install dazzle-dsl
dazzle mcp setup
# Verify
dazzle mcp check| Example | Complexity | What it demonstrates |
|---|---|---|
simple_task |
Beginner | 3 entities, state machine, personas, workspaces, access control |
contact_manager |
Beginner | CRM with relationships and list/detail surfaces |
support_tickets |
Intermediate | Ticket lifecycle with state machines and assignments |
ops_dashboard |
Intermediate | Workspace stages and aggregate metrics |
fieldtest_hub |
Advanced | Full-featured demo with integrations |
pra |
Advanced | 15 DSL files covering every construct: ledgers, processes, LLM, services |
Full LSP implementation: real-time diagnostics, hover docs, go-to-definition, auto-completion, document symbols.
dazzle lsp run # Start the LSP server
dazzle lsp check # Verify dependencies
dazzle lsp grammar-path # TextMate grammar for syntax highlightingWorks with VS Code, Neovim, Emacs, and any editor supporting LSP. See docs/reference/index.md for editor setup.
- Evaluating Dazzle — skeptical-evaluator walkthrough: see the claims demonstrated in ~30 min
- Security & Compliance Claims — claim-by-claim inventory: status, enforcement, tests, known gaps
- Agent Workflow Guide — end-to-end AI-agent spec-edit loop: spec change → DSL edit → validate → tests → human review → deploy
- DSL Reference — complete guide to all DSL constructs
- HLESS deep dive — event semantics and why they're named this way
- Graphs — entity graph relationships, CTE traversal, algorithms
- Compliance — ISO 27001 + SOC 2 evidence pipeline
- RBAC Verification — provable access control
- Enterprise SSO & Provisioning — per-org OIDC / SAML 2.0 / SCIM (opt-in)
- Autonomous Harness — Claude Code slash commands + methodology
- ADRs — architectural decisions, defended
- Architecture — system design, pipeline, MCP server
- Getting Started — installation, quickstart, first app
- Examples — runnable example applications
- Fixtures — framework-validation probes (
shapes_validationfor RBAC,asset_registryforsubtype_of:TPT inheritance)
Dazzle is a research project exploring what application substrates look like when AI collaborators are treated as first-class readers and writers. It is developed in the open, primarily by a single author, with heavy AI assistance — both in the framework itself and in the example apps built on top of it. Release cadence is high (every fix gets a unique version for deployment traceability) and pre-1.0 breaks are intentional rather than apologetic. If you're evaluating Dazzle for production use, talk to us first.
See CONTRIBUTING.md.
MIT — see LICENSE