Intelligence Without Limits

Local AI orchestration with inspectable memory.

ABI is the runtime/WDBX twin under the Abbey/ABI umbrella — receipts, not adjectives. Every surface on this page is backed by code in the repository: capabilities that are partial say so, and capabilities that do not exist are not listed. Cloud backends are optional. Browser autonomy is not Current.

What ships today

Status labels follow the repository's own Current / Partial / Proposed mapping. A Partial label means the surface runs but its scope is narrower than the name suggests — the linked docs spell out the boundary.

CLI Current

Thirteen contract-tested commands behind the abi binary, with golden fixtures pinning the frozen surface. abi help --json emits typed command metadata and shell completions for automation.

WDBX vector store Current

The required sibling implementation has contract coverage for ordered vector search, block metadata round-tripping, segment/WAL recovery and compaction, and temporal graph snapshot restore. Runtime persistence that is skipped or unavailable never reports a fabricated successful write.

MCP server Current

Twelve contract-covered tools over stdio JSON-RPC with a 64 KB frame cap. A custom loopback HTTP compatibility listener is attempted at startup; its /sse response only advertises /message and is not a persistent, spec-conforming MCP HTTP+SSE channel. A bind failure leaves stdio serving.

GPU reporting Honest by design

A capability table and preferred-backend report that returns accelerated=false when native kernels are not linked. Vector ops fall back to deterministic CPU SIMD. CUDA, Vulkan, and ANE residency remain non-claims.

Plugins Current

Sixteen bundled plugins ship as compiled-in mod.rs/stub.rs pairs guarded by assert_plugin_parity!, so a stub can never silently drift from the implementation it stands in for.

Connectors Partial

Discord and Twilio validate credential shape, IDs, payload size, and escaping before dispatch. connector_test exercises deterministic local paths and does not perform live network dispatch.

Agents & scheduler Partial

agent plan, train, multi, and spawn run on the local scheduler-backed helper surface with live task and memory observability. agent browser emits a reviewed local plan only — ABI neither embeds nor launches a browser.

Model runtime Partial

abi-model-runtime requires an exact registry model and explicit device choice. Scratch fixtures prove a tiny abi-bigram-v1 Candle path on CPU plus locally exercised Metal — runtime-foundation evidence, not a quality, placement, or speedup claim.

Workers Partial

abi-worker defines authenticated, audience-bound admission and control contracts with finite leases, cooperative cancellation, replay resistance, and verifiable results — contracts only, with no listener, production cluster, or separate-host deployment claimed.

Quick start

ABI targets nightly Rust pinned by rust-toolchain.toml. Always go through ./tools/cargo.sh — a stable cargo on PATH will shadow the pinned toolchain.

build & validate
# clone the path-dependent repositories as siblings
mkdir abi-workspace
cd abi-workspace
git clone https://github.com/donaldfilimon/wdbx
git clone https://github.com/donaldfilimon/abi
cd abi

# nightly toolchain via the wrapper, never bare cargo
./tools/cargo.sh --version

# the primary gate: fmt, clippy -D warnings, build, tests, docs
./tools/check.sh

# binaries
./tools/cargo.sh build -p abi-cli   # target/debug/abi
./tools/cargo.sh build -p abi-mcp   # target/debug/abi-mcp
local walkthrough
ABI=./target/debug/abi

# capability + scheduler reporting
$ABI backends
$ABI scheduler status
$ABI dashboard --once --plain

# local completion surfaces (no network credentials)
$ABI complete "summarize ABI scheduler status"
$ABI complete --neural "hello"

# agents and the vector store
$ABI agent plan "stage a safe WDBX refactor"
$ABI wdbx stats
$ABI wdbx compute info

MCP smoke test

Build the server and drive the same contract tools from any MCP client. Prefer the launcher (or run from the repository root) so @loader_path resolves the Apple FoundationModels shim next to the binary on arm64 macOS.

stdio server
./tools/cargo.sh build -p abi-mcp
./mcp/launcher.sh

# optional custom loopback HTTP compatibility listener:
#   GET /sse advertises POST /message; it is not persistent MCP SSE
#   ABI_MCP_HTTP_PORT   (default 8080)
#   ABI_MCP_HTTP_TOKEN  (bearer auth)

Live providers are opt-in

complete --live serves Anthropic-provider models over the explicit live transport and therefore needs stored credentials. Model ids are recorded in their catalog-canonical form; an unrecognised id prints a stderr warning and passes through.

credentials
ABI_AUTH_TOKEN=… $ABI auth signin anthropic
$ABI complete --live --model … "hello"

# arm64 macOS with Apple Intelligence ready:
$ABI complete --live --model apple-fm --confirm "hello"

Frozen surfaces

These names are contract-tested against golden fixtures under tests/golden/. Changing one is a deliberate, reviewed act — not a refactor side effect.

CLI commands 13

helpcompletetrain agentbackendsplugin authtwiliotui dashboardwdbxscheduler nn

MCP tools 12

ai_runai_completeai_learn ai_trainwdbx_queryscheduler_stats scheduler_infoconnector_testgpu_status plugin_listwdbx_statsplugin_run

Claims discipline

ABI does not claim production FHE, AES/RBAC, multi-host production deployment, QPS/latency/accuracy numbers, Kubernetes or H100 deployments, CUDA/Vulkan runtime execution, ANE residency, or regulatory certification. Cluster RPC, compute selection, and the security demos are documented honestly as single-host and reference-scoped surfaces.

External collateral — this page included — should cite a repository test, a benchmark artifact, or a documented source file before making a performance or capability claim. The audit that governs this is docs/contracts/external-claims-audit.mdx.