Plan 2.0 - All 30 Phases Complete Β· Public Release 2026-06-21

The AI-Native
Executable Semantic
Protocol

Eve is a local-first runtime designed for AI models to author compact .eve semantic source, compile it into typed DAGs, and generate verified targets - while humans inspect, approve, and govern.

Humans govern.
AI models author.

Eve is an AI-native executable semantic protocol and local-first runtime. AI models write compact .eve source files, which are lowered into a typed semantic DAG, canonical binary artifacts, execution plans, and generated target outputs such as Web applications. Humans inspect, approve, govern, debug, and audit through projections and tools - they are not the primary source authors.

πŸ’‘
Intent
Human expresses intent in natural language
β†’
πŸ“
.eve Source
AI authors compact semantic source
β†’
πŸ”—
Semantic DAG
Typed directed acyclic graph compilation
β†’
βš™οΈ
Plan & Runtime
Execution plan + capability adapters
β†’
🎯
Targets
Generated Web, data, API, native outputs
β†’
βœ…
Receipts & Audit
Replay evidence, audit trail, governance

Semantic truth as source.
Everything else as projection.

Semantic Source of Truth

.eve files are the only truth. Generated traditional code is an output - never a source of logic. Every change flows through the semantic DAG and leaves an auditable trail.

Local-First Runtime

Runs entirely on your machine. No cloud dependency. Semantic DAGs, execution plans, capability RPC, receipts, replay, and checkpoints live under .eve/.

Bounded Capability Model

Every OS interaction - files, network, processes, clipboard, browser - goes through typed capability contracts with allow/deny/approval/budget/redact policies.

Receipts & Replay

Every execution produces cryptographic receipts with SHA-256 content digests. Full replay and checkpoint support for audit, debugging, and compliance.

Web Target Generation

Generate complete HTML/CSS/JS applications from .eve semantic truth. Generated output is fully traceable to its semantic source with target diff reporting.

Incremental Intent Loop

Preview and apply incremental changes via intent fragments. Only affected targets are rebuilt. Digest-anchored patches keep everything traceable.

30
Phases Complete
16
Rust Crates
32K+
Lines of Rust
MIT
Open Source License

One binary, every operation.

The eve binary is the complete toolchain. From source creation to production operations.

eve new my-app --template reference-web

Create a new project from an official template

eve check

Validate and diagnose .eve source files

eve graph

Compile and inspect the semantic DAG

eve plan

Lower DAG into an execution plan

eve run

Execute the plan with capability adapters

eve test

Run semantic tests with capability mocks

eve verify

Verify receipts, replay, and checkpoints

eve build

Generate target artifacts (Web, data, API)

eve doctor

Run full project diagnostics

eve fmt --check

Format and validate .eve source style

eve ops health

Production health check from semantic truth

eve ops replay-incident

Replay and audit production incidents

Plan 2.0 - All 30 phases complete.

Every phase is independently verified, with frozen fixtures, documentation, and retro conclusions.

Phase 1-2
Source & Semantic DAG
.eve parsing, validation, formatting, expansion. Typed semantic DAG compilation and inspection.
Phase 3-4
Canonical Protocol & Execution Engine
Deterministic CBOR encoding, SHA-256 content digests, typed DAG patches. Execution plans, runtime, and checkpoint/replay.
Phase 5-7
Capability RPC & Web Target Generation
Bounded capability contracts, receipt ledger. Project system with eve.toml. Web target generation (HTML/CSS/JS).
Phase 8-12
CLI, Stdlib, Service Mode & Intent Loop
Public command surface. Standard library. Durable service mode with approval. Incremental intent preview/apply loop. Public release declaration.
Phase 13-20
OS Runtime & Desktop Integration
OS capability contracts. Policy engine. Headless adapters. Persistent local agent. Desktop notifications, clipboard, browser automation. Native shell. OS packaging.
Phase 21-27
Package System & Ecosystem
Package manifests. Deterministic lockfiles. Test framework with capability mocks. Debug trace & profiler. LSP conformance. Compatibility policies. Registry.
Phase 28-30
Production Ops & Model-Native GA
Health checks, incident replay, redacted exports. Official template catalog. Model-native GA release with conformance certification and ecosystem governance.

Frequently asked questions.

How is Eve different from traditional programming languages?

Traditional languages are designed for humans to write every line of code. Eve is designed for AI models to author .eve semantic source. The source of truth is always the semantic representation - generated traditional code is an output, never the logic truth. This eliminates the mismatch between what AI models think about (semantics) and what they produce (syntax).

Is the small install size a sign of incompleteness?

No. Eve's binary is intentionally compact - it carries a full parser, semantic DAG compiler, canonical encoder, execution planner, runtime with capability adapters, Web target generator, and production ops toolchain. The small size reflects the protocol's design: semantic truth is expressed compactly, and generated host-code targets are projections, not the runtime itself.

Why does eve build generate traditional code?

The eve build command generates traditional code (HTML/CSS/JS for the Web target) as a projection of the semantic DAG. This provides practical deployment to browsers and existing infrastructure while preserving the .eve source as the sole semantic truth. Every generated file is traceable to its source DAG node, with target-diff reporting for change tracking.

What is the relationship between natural language, AI, Eve, and traditional code?

Natural language β†’ (AI model) β†’ .eve semantic source β†’ (Eve toolchain) β†’ Semantic DAG β†’ Execution Plan β†’ (Eve runtime) β†’ Generated targets & receipts. The AI model handles the intent-to-semantics translation. Eve handles semantics-to-execution. Traditional code is an output surface - browsers, APIs, and native apps consume it, but they never become the truth.

Can any AI model use Eve?

Eve's protocol is model-agnostic. Any AI model that can produce valid .eve source syntax can author Eve projects. The toolchain validates, compiles, and executes the result independently. Phase 30 includes model integration guides and conformance certification for model providers.

Is this a "vibe coding" project?

Yes - the author describes Eve as a personal vibe coding project built in collaboration with GPT, DeepSeek, Qwen, and other AI models. But it has been executed with rigorous engineering discipline: 16 Rust crates, comprehensive tests, frozen fixtures for every phase, and full CI/CD with binary releases.

What platforms and targets does Eve support?

Eve runs on macOS (arm64) and Linux (x86_64). It can generate Web targets (HTML/CSS/JS applications), data workflows, API backends, and local agent automations. Native UI and OS automation are supported through bounded capability adapters with policy-controlled execution.