Silc — Intent-Native Polyglot Application Compiler

Silc (pronounced silk) is ThoughtPivot’s intent language and local Rust compiler for building real software applications with far less accidental complexity — and far fewer model tokens — than asking an LLM to invent a full stack from scratch.

You write a short .silc program that declares what the application is: typed contracts, UI components, resource capabilities, routes, and domain pipelines. The compiler validates that intent, routes each module to the right engine, provisions pinned runtimes, synthesizes dual-surface UI and persistence, and runs a supervised polyglot runtime.

Silc is open source from ThoughtPivot.

.silc intent  →  Rust compiler  →  Bun · CPython · Go workers  →  mmap IPC + UDS

The surface is Raku-inspired, not Raku-compatible. Source files are .silc only. Code examples below use GitHub’s raku fence solely so syntax highlighting renders well.

The problem Silc solves

Modern AI coding workflows still spend most of their budget on decisions that should be deterministic:

Which framework should host the UI?
How should web and terminal surfaces stay in sync?
Where does SQLite wiring live, and who owns migrations?
Which language should score text, call a local LLM, crawl a site, or embed a document?
How do those processes exchange payloads without reinventing glue every time?

Agents and humans repeatedly invent React trees, Python services, Go stores, package manifests, IPC schemes, and deployment scaffolding. That inventiveness burns tokens, creates drift between runs, and blurs the line between product intent and runtime substrate.

Silc’s thesis: authors and agents should declare intent; the compiler should own substrate. Deterministic routing, closed operation registries, and compiler-synthesized mechanics shrink the generation surface. Models spend tokens on domain meaning — forms, inventory, scrapers, assistants — while Silc handles the rest.

Business cases

Silc is for teams that want AI-assisted software creation at application scale, not just snippet scale:

Business need	What Silc provides
Dual-surface internal tools	One component tree → web (React/Tailwind) + terminal (OpenTUI)
Operational CRUD apps	Contracts + resource capabilities → SQLite HTTP APIs
Local AI assistants	Grounded `ui::chat` over live data via silclm
Content / research ingestion	`scrape::*` crawls without naming Bun, Colly, or Playwright; `doc::extract` turns uploads into structured rows
Embedding pipelines	Closed MiniLM path: scrape → tokenize → infer → SQLite
Agent-authored software	Closed language + compiler oracle → fewer invented substrates

The economic pitch is simple: fewer tokens per working application, because framework choice, dual-surface parity, persistence, engine selection, and IPC are compiler decisions — not prompt decisions.

Design principles

Intent over substrate. Authors never write serve(), invent React or OpenTUI trees, declare sinks, or wire ipc::* / store::* pipelines.
Deterministic compilation. Tier 1/2 routing cites engine strengths; every decision has provenance.
Scalable monolith. One cohesive .silc intent model compiles into a supervised cluster of specialized workers (Bun, CPython, Go) that share memory-mapped slots. You author one program; the runtime is polyglot and co-located — not a sprawl of hand-maintained microservices.
AI-native, compiler-first. Models emit .silc. The compiler is the validation oracle. Assist explores corpus and checks drafts without stuffing the entire authoring contract into the root prompt.
Pinned, owned runtimes. Bun, CPython, and Go are checksum-verified into ~/.silc/runtimes/. Authors and agents do not choose engines.

How it looks in practice

Examples below are Silc 0.4.0 source. GitHub fences use raku for highlighting only.

1. A dual-surface notes app

What silc init scaffolds — a form, an app route table, and an optional scorer. Dual-surface web/terminal serving and SQLite persistence are synthesized.

@version("0.4.0")

contract Note {
    has Str $.author;
    has Str $.text;
}

component HomePage {
    has state Str $.author = "";
    has state Str $.text = "";

    method render() {
        ui::page(
            :app_bar(ui::app_bar(:title("My Silc App"))),
            :side_panel(ui::side_panel(
                ui::nav_item(:label("Home"), :to("/"), :active)
            )),
            ui::stack(
                ui::heading(:text("Leave a note"), :level(2)),
                ui::form(:on(submit(on_submit)),
                    ui::text_input(:field(author), :label("Author")),
                    ui::textarea(:field(text), :label("Note")),
                    ui::toolbar(
                        ui::button(:label("Submit"), :variant(primary), :submit)
                    )
                )
            )
        )
    }

    method on_submit() {
        submit();
    }
}

app MyApp {
    route "/" => HomePage;
}

processor NoteScorer {
    method analyze(Note $note) {
        $note.text ==> text::score()
    }
}

You declared: schema, UI, routes, scoring intent. Silc synthesizes: React web + OpenTUI terminal, POST /submit, Go/SQLite sink, Bun ingress, and mmap staging between workers.

2. Resource CRUD + grounded local chat

From examples/inventoryApp — capability-style resources become HTTP CRUD; chat is grounded on a live inventory snapshot.

contract InventoryItem {
    has Str $.id;
    has Str $.name;
    has Str $.category;
    has Str $.location;
    has Str $.quantity;
    has Str $.reorder_level;
    has Str $.notes;
}

contract ChatRecord {
    has Str $.prompt;
    has Str $.reply;
}

resource InventoryItems for InventoryItem {
    query list;
    mutation create;
    mutation update;
    mutation delete;
}

component BrowsePage {
    has state Str $.category_filter = "All";
    query $.items = InventoryItems.list();

    method render() {
        ui::page(
            :app_bar(ui::app_bar(:title("Inventory"))),
            :side_panel(ui::side_panel(
                ui::nav_item(:label("Browse"), :to("/"), :active),
                ui::nav_item(:label("Admin"), :to("/admin")),
                ui::nav_item(:label("Assistant"), :to("/assistant"))
            )),
            ui::stack(
                ui::section(
                    :title("Stock browser"),
                    :description("Filter by category, or ask the Assistant about live inventory.")
                ),
                ui::table(
                    :rows($.items),
                    :columns(["name", "category", "location", "quantity", "reorder_level", "notes"]),
                    :empty_text("No inventory items yet. Add some in Admin."),
                    :filter_field(category_filter),
                    :filter_column("category"),
                    :sortable,
                    :searchable
                )
            )
        )
    }
}

# … AdminPage omitted …

component AssistantPage {
    has state Str $.prompt = "";
    query $.items = InventoryItems.list();

    method render() {
        ui::page(
            :app_bar(ui::app_bar(:title("Inventory Assistant"))),
            ui::chat(
                :value($.prompt),
                :context($.items),
                :persona("You are the Inventory Assistant for this Silc inventory app, built on silclm."),
                :placeholder("Which items are below reorder level?"),
                :on(send(on_send))
            )
        )
    }

    method on_send() {
        Assistant.complete();
    }
}

app InventoryApp {
    route "/" => BrowsePage;
    route "/admin" => AdminPage;
    route "/assistant" => AssistantPage;
}

processor Assistant {
    method complete(ChatRecord $record) {
        $record.prompt ==> llm::complete()
    }
}

You declared: domain model, CRUD capabilities, browse/admin/assistant routes, and a local completion processor. Silc synthesizes: /api/inventory_items CRUD, dual-surface UI, silclm provisioning, and persistence for chat/processor results.

3. Pipeline-only: scrape → embed → store

From examples/pipelineApp — no UI app required. One intent file becomes a Bun/CPython/Go ingestion graph.

@version("0.4.0")

subset Uri of Str where { .starts-with("http") }
subset Emb384 of Vec[num32; 384];

contract ArticlePayload {
    has UUID $.id;
    has Uri $.url;
    has Str $.raw_content;
    has Emb384 $.vector_embedding;
}

service ArticleIngress {
    method fetch_article() {
        target_url
            ==> scrape::page(:js(false))
            ==> scrape::extract(:into(ArticlePayload))
    }
}

processor Embedder {
    method embed(ArticlePayload $article) {
        $article.raw_content
            ==> tensor::tokenize(:model("minilm-l6-v2"))
            ==> tensor::infer(:prefer(CPU))
    }
}

Run with:

silc run main.silc --input-json '{"url":"https://example.com/"}'

Runtime: why Bun, CPython, and Go

Silc does not ask models (or developers) to pick languages. The router assigns work from complementary strengths (ADR-004):

Engine	Role in Silc
Bun	Generated TypeScript: web UI, terminal UI, HTTP ingress, static scrape helpers
CPython	Scoring, local LLM (llama.cpp / silclm), Playwright scrape, ONNX MiniLM
Go	SQLite persistence, HTTP APIs, high-concurrency Colly crawls

Engines are pinned and checksum-verified (Bun 1.2.18, CPython 3.12.12, Go 1.23.6) under ~/.silc/runtimes/. There is no PATH override surface and no author-facing engine picker.

Silc source (.silc)
        │
        ▼
   sil-lexer → sil-parser → sil-core subjects
        │     (Contract · Component · Resource · App · Module · Pipeline)
        ▼
   sil-router   Tier 1 (kind + traits) + Tier 2 (namespaces)
        ▼
   sil-codegen  runnable workers + dual-surface UI lowering
        ▼
   silc supervisor
        ├── Bun  (web + terminal + resource HTTP + static scrape)
        ├── CPython (scoring / local LLM / Playwright / ONNX)
        ├── Go (SQLite / HTTP API / Colly crawl)
        └── sil-ipc mmap slots + UDS

Scalable monolith

A Silc program is a monolith at the intent layer and a supervised polyglot runtime underneath. One file owns the product model. The compiler emits specialized workers that scale within that model — for example, replica pools for CPU-bound scoring — without forcing authors to design a microservice mesh. That is the scalable-monolith shape: cohesive product semantics, partitioned execution, shared contracts.

IPC that stays out of your way

Cross-engine data movement uses ThoughtPivot’s Silc Shared Buffer ABI v1 (ADR-001, SILC-IPC-ABI-v1.md):

Data plane: file-backed mmap slots under .runtime/ (default 512 × 16 KiB; larger for pipeline payloads). Magic bytes SILC.
Control plane: small Unix domain socket wakeups (segment_id, offset, len, schema_id).

Payloads stay in shared memory between processor and synthesized persistence. Workers do not retransmit application bodies over HTTP between those stages. ABI v1 carries schema-tagged JSON in the mapped buffer; typed zero-copy field views are a future ABI layer, not a current claim.

AI-native authoring and Silc Assist

Silc is designed so language models author intent programs, not framework scaffolding.

In-app intelligence: llm::complete / ui::chat run on silclm (compiler-pinned local GGUF). Use :context(...) to ground answers on live resource data.
Silc Assist (experimental): silc assist drafts and modifies .silc files with silclm (ADR-008). It auto-retrieves relevant examples and AGENTS.md rules, asks for a complete program via the chat template (stop marker # END), then compile-and-repairs. Creating a file adapts the silc init starter as a skeleton, so the usual run lands on the first attempt in ~6–12s. Repairs escalate cheapest-first: mechanical diagnostics (a repeated resource block, a resource named like a component, seeds missing :id, a nested method, a missing @version) are auto-fixed with no model call, structural ones get an explicit rule, and only the rest fall back to error-targeted corpus search. Tasks that ask to persist data get the resource + mutation pattern injected up front. An identical repeated draft retries at a higher temperature instead of recompiling the same file, and if every attempt fails the closest draft is saved as <file>.rejected for inspection. The slower tool loop is opt-in (--explore). While it runs, the terminal shows a durable action trace plus a spinner — not raw model protocol. Inference uses a warm silclm worker with Metal GPU offload by default on Apple Silicon.

silc assist "dual-surface notes app with submit" notes.silc
silc assist "refine the form" notes.silc --explore   # optional slower fallback

Assist is Phase 1: useful, bounded, and experimental. A fine-tuned silclm-assist model is reserved but not shipped yet. In-app chat and Assist remain separate products on the same local model family.

Token efficiency, concretely: every framework/engine/persistence decision the compiler owns is a decision the model no longer has to negotiate in context. Compiler diagnostics then act as a hard oracle — accepted programs parse, validate, and route before they run.

Quick start

cargo install --path crates/silc --force

silc init myapp
cd myapp
silc build main.silc   # validate + codegen
silc main.silc              # run web by default
silc main.silc --terminal   # also attach OpenTUI (+ telnet)

# web:      http://127.0.0.1:18088  (override SILC_HTTP_PORT)
# terminal: silc main.silc --terminal  (or SILC_TERMINAL=1)
# fallback: telnet 127.0.0.1 18023 when --terminal is set

silc init writes main.silc, AGENTS.md, .gitignore, and a runtime lock, then provisions pinned engines on first use.

Example projects

App	Purpose
`examples/chatApp/`	Multi-session local chat via silclm
`examples/inventoryApp/`	CRUD + browse/admin + grounded assistant
`examples/scraperApp/`	URL + depth crawl; results table + summaries
`examples/pipelineApp/`	Scrape → MiniLM/ONNX → SQLite
`examples/blogApp/`	Seeded blog; year/month filters; admin modal CRUD; grounded search

See examples/README.md.

Editor support (VS Code / Cursor)

Silc ships a VS Code / Cursor extension that provides syntax highlighting and a Rust language server (sil-lsp) for semantic hover on .silc sources — resource methods, query bindings, contracts and fields, components, props and state, UI primitives, executable ops, keywords, operators, and builtin types.

Install it with the bundled script:

./editors/vscode-silc/install.sh

The script:

Builds sil-lsp in release mode (cargo build -p sil-lsp --release)
Installs npm dependencies and compiles the TypeScript language client
Bundles the host-platform server binary into a VSIX
Installs the extension with the cursor CLI, falling back to code

Requirements: a Rust toolchain, Node.js/npm, and a cursor (or code) CLI on your PATH. In Cursor, you can add the CLI via Shell Command: Install 'cursor' command in PATH. Set SILC_EDITOR_CLI to override CLI detection.

After it finishes, run Developer: Reload Window. Open any .silc file — the language indicator should read Silc, and hovering a symbol should show a Markdown tooltip. To point the editor at a locally built server without reinstalling, set silc.languageServerPath to your target/release/sil-lsp path.

See editors/vscode-silc/README.md for hover coverage, highlighting scopes, and development details.

What ships today (0.4.0)

Silc is pre-1.0. Release 0.4.0 makes the product rule explicit: authors declare intent; the compiler synthesizes runtime mechanics (ADR-009).

Every UI app synthesizes both surfaces automatically — compiler-owned ui::web (React/Tailwind) and ui::terminal (OpenTUI). Authors declare routes only; they never write method serve(), ui::web, or ui::terminal as program operations. The full UI primitive catalog (39 dual-surface builtins), closed prop enums, and agent rules live in crates/silc/templates/AGENTS.md.

Executable operations

Author-facing ops that run today:

service::http, text::score, llm::complete, scrape::page, scrape::site, scrape::select, scrape::render, scrape::extract, doc::extract, tensor::tokenize, tensor::infer.

Shipped

Parse → validate → deterministic Tier 1/2 route → codegen → supervised run
Declaration-based component / resource Name for Contract / app routes
Dual-surface UI synthesized from app (web + terminal)
Generic resource CRUD over SQLite
silc init scaffold and experimental silc assist
Compiler-owned Bun / CPython / Go under ~/.silc/runtimes/

Boundaries

Broader pipeline namespaces (http::*, html::*, numpy::*, pandas::*, …) are stub-only: they parse/route/emit but do not execute
Tensor path is CPU-only MiniLM → exactly 384 normalized num32 values
IPC ABI v1 is schema-tagged JSON in mmap (not typed zero-copy views)
No self-contained silc bundle deployment artifact yet
Assist is experimental; fine-tuned assist weights are not shipped

Authoring contract for agents: crates/silc/templates/AGENTS.md.

For AI agents

silc init copies the agent contract into the project:

Edit .silc only — never patch .runtime/
Declare routes; dual-surface serving is synthesized
Prefer components + resources over inventing portal profiles or frameworks
Stay inside the UI catalog and runnable operation set
Validate with silc build; report limits instead of escaping to React/OpenTUI

Development

cargo fmt --all -- --check
cargo check --workspace
cargo test --workspace -- --test-threads=1

CI runs fmt, check, library tests, codegen smoke, dual-surface e2e builds, and concurrent /submit POSTs with SQLite checks.

Versioning

Pre-1.0 SemVer 0.x: breaking language/compiler changes bump the minor. 1.0.0 is reserved for a future stability milestone. Releases use release-plz and Conventional Commits.

Documentation

Doc	Topic
docs/ADR-INDEX.md	Decision index
docs/ARCHITECTURE.md	Subject model and crate layout
docs/intent-vs-subjects.md	Intent authoring vs subject architecture
docs/ADR-001-runtime-and-ipc.md	Engines and IPC
docs/ADR-002-silc-surface-syntax.md	Language surface
docs/ADR-003-declarative-ui.md	Dual-surface UI policy
docs/ADR-004-runtime-strengths.md	Why Bun / CPython / Go
docs/ADR-005-local-llm-complete.md	Local LLM completions
docs/ADR-006-scrape-namespace.md	`scrape::*`
docs/ADR-007-pipeline-feeds.md	`==>` semantics
docs/ADR-008-recursive-silclm-assist.md	Silc Assist
docs/ADR-009-compiler-synthesized-runtime.md	Synthesized UI / persistence
docs/ADR-010-tensor-minilm-pipeline.md	MiniLM embedding pipeline
docs/ADR-011-document-extract.md	`doc::*` upload + extract
docs/SILC-IPC-ABI-v1.md	Shared buffer ABI
CHANGELOG.md	Release notes

License

Apache-2.0 — see LICENSE.

Maintained by the ThoughtPivot engineering team.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Repository files navigation

Silc — Intent-Native Polyglot Application Compiler

The problem Silc solves

Business cases

Design principles

How it looks in practice

1. A dual-surface notes app

2. Resource CRUD + grounded local chat

3. Pipeline-only: scrape → embed → store

Runtime: why Bun, CPython, and Go

Scalable monolith

IPC that stays out of your way

AI-native authoring and Silc Assist

Quick start

Example projects

Editor support (VS Code / Cursor)

What ships today (0.4.0)

Executable operations

For AI agents

Development

Versioning

Documentation

License

About

Releases

Packages

Contributors

Languages

Name		Name	Last commit message	Last commit date
Latest commit History 51 Commits
.github/workflows		.github/workflows
assets/brand		assets/brand
crates		crates
docs		docs
editors/vscode-silc		editors/vscode-silc
examples		examples
training		training
.gitignore		.gitignore
CHANGELOG.md		CHANGELOG.md
Cargo.lock		Cargo.lock
Cargo.toml		Cargo.toml
LICENSE		LICENSE
README.md		README.md
release-plz.toml		release-plz.toml
rust-toolchain.toml		rust-toolchain.toml

Uh oh!

Folders and files

Latest commit

History

Repository files navigation

Silc — Intent-Native Polyglot Application Compiler

The problem Silc solves

Business cases

Design principles

How it looks in practice

1. A dual-surface notes app

2. Resource CRUD + grounded local chat

3. Pipeline-only: scrape → embed → store

Runtime: why Bun, CPython, and Go

Scalable monolith

IPC that stays out of your way

AI-native authoring and Silc Assist

Quick start

Example projects

Editor support (VS Code / Cursor)

What ships today (0.4.0)

Executable operations

For AI agents

Development

Versioning

Documentation

License

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages