sbhooley/ainativelang

pipx install 'ainativelang[mcp]' && ainl setup --auto  (one-command agent install, auto-detects hosts)
pip install ainativelang  (CLI only)
pip install -e ".[mcp]" && ainl-mcp  (MCP server, stdio)
ainl install-mcp --host openclaw|hermes|zeroclaw|armaraos  (per-host install)

Usage

• •Create a project: ainl init my-first-worker, then ainl check main.ainl --strict, ainl run main.ainl, ainl visualize main.ainl --output - (Mermaid).
• •Author in compact syntax (recommended, 66% fewer tokens) or opcode syntax; both compile to the same IR.
• •Emit to other platforms: ainl emit your.ainl --target langgraph -o graph.py, or serve as HTTP API with ainl serve --port 8080.
• •From an MCP host, call ainl_validate, ainl_compile, ainl_capabilities, ainl_security_report, and ainl_run over the ainl-mcp stdio server.

Key features

• ✓Graph-first deterministic execution: compile-once, run-many; no LLM re-generation for graph routing and adapter orchestration on each run (claimed ~90-95% fewer orchestration tokens vs prompt loops on recurring monitors).
• ✓Strict-mode guarantees: canonical graph emission, no undeclared references, adapter arity validation, controlled exits, with structured diagnostics and llm_repair_hint.
• ✓Adapter-based effect system: 29 core + 16 extended adapters (http, sqlite, postgres, mysql, redis, dynamodb, airtable, supabase, fs, tools, a2a, code_context, etc.) with privilege tiers and security profiles.
• ✓Multi-target emission: FastAPI, React/TypeScript, Prisma, OpenAPI, Docker/Compose, Kubernetes, cron/queue/scraper, LangGraph and Temporal hybrid wrappers.
• ✓MCP v1 server with safe defaults (core-only adapters, conservative limits), exposure profiles, capability grants, per-run adapters, and workspace limits.

Best for

Teams running 20+ recurring monitor/digest/scheduled workflows driven by LLM prompt loops who want deterministic, auditable, compile-once-run-many orchestration with strict validation.

Caveats

• ⚠Token savings only apply vs prompt-loop baselines (~90-95% if re-prompting every run, ~1.3-1.5x vs hand-optimized runners, ~0% vs pure deterministic runners).
• ⚠Open-core: enterprise audit/policy packs, managed ops, and hosted runtime are paid/optional; core DSL, compiler, runtime, and MCP server are open (Apache-2.0).
• ⚠No third-party (Class a) paying customer deployment publicly committed yet; evidence is operator deployments and modeled benchmarks.
• ⚠AINL is the workflow layer, not a sandbox/security layer - containment and network policy are the hosting environment's responsibility.
• ⚠Requires Python 3.10+; some CLI features (ainl dashboard) fail fast if the emitted demo server is missing on PyPI-only installs.