ABI

Orchestration you can inspect on your machine.

ABI routes requests, assembles context, and coordinates tools. The public GitHub tree is nightly Rust. The Zig tree has been removed. WDBX is a required sibling. Follow the repository README — do not mix toolchains. Local template completion does not establish model quality.

CurrentRequires the sibling WDBX workspace.

Responsibilities

What ABI actually does.

Model routing

Partial

Exact registry model IDs and explicit device choice. Optional live providers behind stored credentials.

Context management

Current

Inspectable request context before a model or tool runs. Persistence records vectors and metadata only when the write succeeds.

Safety boundaries

Partial

Claim-honest reporting, plugin parity guards, bounded workers with finite leases and replay resistance (contracts).

Tool orchestration

Current

Twelve MCP tools in the public tree over stdio, plus an optional loopback HTTP+SSE listener. Not Streamable HTTP.

Reasoning coordination

Partial

Scheduler-backed helpers with live task and memory observability where the tree implements them.

Memory integration

Current

WDBX is a required sibling. Skipping persistence never fabricates a successful write.

Public tree

Nightly Rust. Use the repo wrappers.

Bare cargo is the wrong entry. ./tools/cargo.sh pins the toolchain the tree actually builds with.

donaldfilimon/abi
git clone https://github.com/donaldfilimon/wdbx
git clone https://github.com/donaldfilimon/abi
cd abi
./tools/cargo.sh
./tools/check.sh
./tools/cargo.sh build -p abi-cli

GPU reporting returns accelerated=false when native kernels are not linked. Persistence defaults to $HOME/.abi/wdbx. Disable with ABI_WDBX_PERSIST=0. License: Apache-2.0.

Crates

What the workspace actually ships.

  • abi-cli

    Operator surface. backends, scheduler, dashboard, plugin, wdbx.

  • abi-mcp

    JSON-RPC 2.0 over stdio, optional loopback HTTP with bearer auth.

  • abi-ai

    Exact model registry and template completion. Quality is not inferred.

  • abi-sea

    Bounded evidence selection and scoring. The learning loop persists router weights.

  • abi-gpu

    Capability reporting. accelerated=false when native kernels are not linked.

MCP

Contract-covered tools, not a hosted SDK.

stdio is the default. The optional loopback listener serves MCP 2024-11-05 HTTP+SSE sessions with bearer auth; it is not Streamable HTTP. Tool names are from the ABI README.

  • ai_learn

    Learning entry point with local side effects. Persistence can be disabled.

  • ai_complete

    Local persona-template completion. The model id is metadata, not proof it ran.

  • wdbx_query

    Nearest-neighbor retrieval with inspectable hits.

  • wdbx_stats

    Local store statistics. Not a cluster dashboard.

  • gpu_status

    Honest device report. Fallback is reported as fallback.

  • plugin_list

    Contract-covered plugins visible to this process.

CLI

Commands the public tree actually documents.

Copied from donaldfilimon/abi. Local template completion does not establish model quality.

  • abi backends

    List configured execution backends and what they actually report.

  • abi scheduler status

    Scheduler health for this process. Not a fleet view.

  • abi dashboard --once --plain

    One-shot text dashboard. No hosted UI implied.

  • abi plugin list

    Plugins the current binary loaded under contract.

  • abi wdbx query

    Retrieve from the local store. Requires the sibling workspace.

abi-cli
$ABI backends
$ABI scheduler status
$ABI dashboard --once --plain
$ABI complete "summarize ABI scheduler status"
$ABI agent plan "stage a safe WDBX refactor"
$ABI help --json

External claims

Do not cite what the ledger does not prove.

External collateral should cite a repository test, a benchmark artifact, or a documented source file before making a performance or capability claim.

  • A current Zig public tree
  • Distributed production sharding
  • AES/RBAC as a complete product
  • Python/TensorFlow stacks
  • Kubernetes or H100 deployments
  • QPS, latency, or accuracy numbers
  • Energy-efficiency comparisons
  • Browser autonomy