Abi
ABI Agent + WDBX Database: A high-speed, fully optimized AI and ML training stack developed in Zig 0.16.0-dev, offering exceptional performance and reliability.
Install / Use
/learn @donaldfilimon/AbiREADME
ABI Framework
ABI is a Zig 0.16 framework for AI services, semantic vector storage, GPU acceleration,
and distributed runtime. The public package entrypoint is src/root.zig, exposed to
consumers as @import("abi").
What ABI includes
abi.App/abi.AppBuilderfor framework setup and feature wiringabi.databasefor the semantic store and vector search surfaceabi.aifor agents, profiles, training, reasoning, and LLM supportabi.inferencefor engine, scheduler, sampler, and paged KV cache primitivesabi.gpu/abi.Gpu/abi.GpuBackendfor unified compute backendsabi.runtime,abi.platform,abi.connectors,abi.mcp,abi.acp,abi.tasksfor servicesabi.foundationfor shared utilities (SIMD, logging, security, time)
Quick start
git clone https://github.com/donaldfilimon/abi.git
cd abi
./build.sh # macOS 26.4+ (auto-relinks with Apple ld)
zig build # Linux / older macOS
On macOS 26.4+ (Darwin 25.x), stock prebuilt Zig's LLD linker cannot link binaries.
Use ./build.sh which auto-relinks with Apple's native linker. To install and symlink
the correct Zig version:
tools/zigly --status # Auto-install if missing
tools/zigly --link # Symlink to ~/.local/bin
Build commands
./build.sh # Build (macOS 26.4+ auto-relinks with Apple ld)
./build.sh test --summary all # Run tests via wrapper (macOS 26.4+)
zig build # Build static library (Linux / older macOS)
zig build test --summary all # Run tests
zig build check # Lint + test + stub parity
zig build lint # Check formatting
zig build fix # Auto-format
zig build check-parity # Verify mod/stub declaration parity
zig build feature-tests # Run feature integration and parity tests
zig build mcp-tests # Run MCP integration tests
zig build messaging-tests # Run messaging unit + integration tests
zig build secrets-tests # Run secrets unit + integration tests
zig build pitr-tests # Run PITR unit + integration tests
zig build agents-tests # Run agents unit + integration tests
zig build multi-agent-tests # Run multi-agent unit + integration tests
zig build orchestration-tests # Run orchestration unit + integration tests
zig build gateway-tests # Run gateway unit + integration tests
zig build inference-tests # Run inference unit + integration tests
zig build cli-tests # Run CLI tests
zig build tui-tests # Run TUI tests
zig build typecheck # Compile-only validation for the current/selected target
zig build validate-flags # Validate feature flags
zig build full-check # Run full check
zig build verify-all # Verify all components
zig build cross-check # Verify cross-compilation (linux, wasi, x86_64)
zig build lib # Build static library artifact
zig build mcp # Build MCP stdio server (zig-out/bin/abi-mcp)
zig build doctor # Report build configuration and diagnostics
CLI
Build the CLI binary with zig build cli (or ./build.sh cli on macOS 26.4+).
The binary lands at zig-out/bin/abi. Entry point: src/main.zig.
abi # Smart status (feature count + available commands)
abi version # Print version and build info
abi doctor # Run diagnostics (all feature flags + GPU backends)
abi features # List all 35 features with [+]/[-] status
abi platform # Show platform detection (OS, arch, CPU, GPU)
abi connectors # List 16 LLM provider connectors
abi info # Framework architecture summary
abi chat <message...> # Route through multi-profile pipeline
abi db <cmd> # Vector database (add, query, stats, diagnostics, optimize, backup, restore, serve)
abi dashboard # Interactive TUI (requires -Dfeat-tui=true)
abi help # Full help reference
Pipeline architecture
The multi-profile pipeline (Abbey-Aviva-Abi) is wired end-to-end in src/features/ai/profile/router.zig:
User Input
-> Abi Analysis (sentiment + policy + rules)
-> Modulation (EMA user preference learning)
-> Routing Decision (single / parallel / consensus)
-> Profile Execution (Abbey / Aviva / Abi)
-> Constitution (6-principle ethical validation)
-> WDBX Memory (cryptographic block-chain storage)
-> Response
Feature flags
All features default to enabled except feat-mobile and feat-tui (both false). Disable with -Dfeat-<name>=false:
zig build -Dfeat-gpu=false -Dfeat-ai=false # Disable GPU and AI features
zig build -Dfeat-training=false # Disable training sub-feature only
zig build -Dfeat-mobile=true # Enable mobile (off by default)
zig build -Dgpu-backend=metal # Single GPU backend
zig build -Dgpu-backend=cuda,vulkan # Multiple GPU backends
Public surface
| Surface | Purpose |
|---------|---------|
| abi.App / abi.AppBuilder | Framework lifecycle and feature orchestration |
| abi.database | Semantic store, search, backup, restore, diagnostics |
| abi.ai | Agents, profiles, LLM, training, reasoning |
| abi.inference | Engine, scheduler, sampler, and paged KV cache primitives |
| abi.gpu / abi.Gpu / abi.GpuBackend | GPU compute backends |
| abi.foundation / abi.runtime | Shared foundations, time/sync/SIMD, runtime primitives |
| abi.search | BM25 full-text search with inverted index persistence |
| abi.connectors / abi.ha / abi.tasks / abi.lsp / abi.mcp / abi.acp | Service and integration surfaces |
Project structure
abi/
├── src/ # Framework source (single "abi" module)
│ ├── root.zig # Public package entrypoint (@import("abi"))
│ ├── core/ # Always-on framework internals
│ ├── features/ # 20 feature directories (35 features including AI sub-features)
│ ├── foundation/ # Shared utilities: logging, security, time, SIMD, sync
│ ├── runtime/ # Task scheduling, event loops, concurrency
│ ├── platform/ # OS detection, capabilities, environment
│ ├── connectors/ # External service adapters (OpenAI, Anthropic, Discord)
│ ├── protocols/ # MCP, LSP, ACP, HA protocol implementations
│ ├── tasks/ # Task management, async job queues
│ └── inference/ # ML inference: engine, scheduler, sampler, KV cache
├── build.zig # Self-contained build root
├── build.zig.zon # Package manifest
└── tools/ # zigly, crossbuild.sh, auto_update.sh
Each feature under src/features/<name>/ follows the mod/stub contract:
mod.zig (real implementation), stub.zig (API-compatible no-ops), and types.zig
(shared types). The stub provides the same public declarations so code compiles
regardless of which features are enabled.
Tools
tools/zigly --status # Print zig path (auto-install if missing)
tools/zigly --install # Force re-download zig + ZLS
tools/zigly --link # Symlink zig + zls into ~/.local/bin
tools/zigly --unlink # Remove symlinks from local bin
tools/zigly --update # Check for newer zig and update if available
tools/zigly --check # Report if update available (no download)
tools/zigly --clean # Remove all cached versions
tools/crossbuild.sh # Cross-compile for linux, wasi, x86_64 targets
tools/auto_update.sh # Check and apply updates for zig + zls
Cache location: ~/.zigly/versions/<version>/bin/{zig,zls}
Toolchain
ABI is pinned to the Zig version in .zigversion. On macOS 26.4+, ./build.sh
auto-relinks with Apple's native linker. On Linux / older macOS, zig build works directly.
Testing
zig build test --summary all # Run all unit + integration tests
./build.sh test --summary all # Same, via macOS 26.4+ wrapper
zig build check # Full gate: lint + test + stub parity
zig build check-parity # Verify mod/stub declaration parity only
zig build messaging-tests # Focused messaging runtime lane
zig build secrets-tests # Focused secrets runtime lane
zig build pitr-tests # Focused PITR runtime lane
zig build agents-tests # Focused agents runtime lane
zig build multi-agent-tests # Focused multi-agent runtime lane
zig build orchestration-tests # Focused orchestration runtime lane
zig build gateway-tests # Focused gateway runtime lane
zig build inference-tests # Focused inference runtime lane
Two test suites run under zig build test:
- Unit tests (
src/root.zig) --refAllDeclswalks the entire module tree. - Integration tests (
test/mod.zig) -- 50 modules covering database, inference, profile pipeline, security, CLI, TUI, and all features.
Specification
See docs/spec/ABBEY-SPEC.md for the comprehensive mega spec covering architecture, profiles, behavioral model, math foundations, ethics, and benchmarks.
License
See LICENSE.
