AWF is a professional ecosystem designed to give engineers full control over their AI integration. From local orchestration to global production.
AWF CLI orchestrates your AI workflows. ZPM gives your agents a memory they can reason over. ztick fires them at the precise moment you decide. Use one, use all three — they're designed to work together.
Go-powered, provider-agnostic, YAML-native. Define workflows once, run them from your terminal, wire them into CI. The foundation everyone builds on.
Zig + Prolog. Give your LLMs deterministic logical reasoning. 22 MCP tools for facts, rules, queries, proofs, and truth maintenance — usable as MCP server or direct CLI.
Zig-powered, push-driven scheduler. Cron in reverse: your apps send SET over TCP at the moment they decide, ztick fires at the requested time — into shells, webhooks, AMQP, Redis, or AWF workflows.
Vendor-agnostic at the orchestration layer — OpenAI, Claude, Gemini, Ollama — and protocol-open at the reasoning layer via MCP. Zero lock-in, OSS end-to-end.
CLI workflows replay identically on the same inputs; ZPM deductions are provable through Prolog inference chains. AI becomes as predictable as the rest of your stack.
Go and Zig, two static binaries, zero runtime dependencies. Structured logs, audit trails, and proof explanations across both products. Built for enterprise, not demos.
Concrete problems the ecosystem solves, in plain terms.
Workflows in YAML, one static binary, zero SDK sprawl. Install, write, run — today.
Every token tracked per step, per workflow, per run. Audit spend line by line — catch runaway agents before they run away.
OpenAI raising prices? Switch to Claude. Need privacy? Drop in Ollama. Edit one line of YAML — your workflows don't care.
CLI workflows replay identically. ZPM proves each conclusion through explicit inference chains. "Why did the AI say X?" finally has an answer.
Give LLMs a knowledge base that reasons. Facts persist, rules apply, contradictions surface. "The AI hallucinated" becomes "the assumption was wrong".
Two static binaries, zero runtime dependencies. Docker if you want, bare metal if you prefer. Any CI, any server, any laptop.
The reasons you'll actually end up installing it tonight.
The foundation of the ecosystem. A lightweight, powerful tool to build and run AI workflows from your terminal.
Written in Go for blazing performance. Every millisecond counts in production.
Switch between OpenAI, Anthropic, or local LLMs without changing a line of code.
Define workflows in simple, version-controllable files. Human readable, machine precise.
Built-in validation and error handling for mission-critical deployments.
# Code review workflow
name: code-review
version: "1.0.0"
inputs:
- name: file
type: string
required: true
states:
initial: read
read:
type: step
command: cat "{{.inputs.file}}"
on_success: analyze
analyze:
type: agent
provider: claude
prompt: |
Review this code for bugs and security issues:
{{.states.read.Output}}
on_success: report
report:
type: step
command: echo "Analysis complete"
on_success: done
done:
type: terminal
status: successDefine your entire workflow in a single YAML file. Human-readable, version-controllable, and production-ready.
Iterate over arrays, file lists, or API results with full workflow context inside each iteration.
type: foreach
Native session resume on Claude, Codex, Gemini, OpenCode. Token tracking and context injection across turns.
type: conversation
Every run has an ID. Interrupt it, fix it, resume from the failed step — no need to start over.
awf resume <id>
Call one workflow from another with typed input/output mapping. Build libraries of reusable pipelines.
type: call_workflow
Run multiple steps concurrently with configurable strategies — all, any, race. Collect results in one place.
type: parallel
Exponential, linear, or constant backoff per step. Flaky APIs and rate limits stop waking you up at 3 a.m.
retry: {strategy: exponential}
Generate Graphviz DOT for any workflow. Render it, ship it to architecture docs, impress the auditor.
awf diagram <workflow>
Paired start/end entries per step, atomic writes, secret redaction, replayable history. Compliance happy.
awf history
Execute your workflows with a single command. AWF provides detailed logging, step tracking, and clear completion status with workflow IDs for reproducibility.
$ awf run code-review --input file=service.go --output=streaming --verbose=true
registered desktop notification backend
registered webhook notification backend
Running workflow: code-review
starting workflow workflow=code-review id=3eb274fa-186e-4070-a6a0-c26eea61ad28
executing step step=read
executing step step=analyze
executing agent step step=analyze provider=claude
executing step step=report
[OUT] Analysis complete
recorded execution history workflow_id=3eb274fa-186e-4070-a6a0-c26eea61ad28 status=completed
workflow completed step=done status=completed
✓ read: completed successfully
Workflow completed successfully in 36.6s
Workflow ID: 3eb274fa-186e-4070-a6a0-c26eea61ad28
--- Execution Details ---
Status: completed
Steps executed:
report: completed (1ms)
read: completed (1ms)
analyze: completed (36.585s)AWF provides detailed execution logs, validation feedback, and clear success/failure status for every workflow run.
The core engine of the AWF ecosystem. Install it in seconds and start building.
curl -fsSL https://raw.githubusercontent.com/awf-project/cli/main/scripts/install.sh | shgo install github.com/awf-project/cli/cmd/awf@latestgit clone https://github.com/awf-project/cli.git
cd cli
make build
make installThe AWF CLI isn't just another walled garden. It leverages the tools and AI providers you already use and trust.
OpenAI
Claude
Gemini
OpenCode
Ollama
Bash
Git
Docker
PHP
Python
Slack
Jira
If it's in your PATH, it's in your workflow. The CLI orchestrates any CLI utility with zero configuration.
The CLI is a platform, not just a tool. Write your own step types or share a whole workflow as a reusable pack.
Plugins add new primitives to the workflow DSL — inject real-time data, call internal APIs, run custom tools. Distribute them as binaries, load them from anywhere.
Packs bundle a complete YAML workflow with its prompts and helpers. Install a pack, run it, or fork it as a starting point for yours.
Built your own? Tag your repo awf-plugin or awf-workflow on GitHub to get discovered.
The reasoning layer of the ecosystem. A Zig-powered MCP server that gives LLMs deterministic logical reasoning through an embedded Prolog engine.
Trealla-powered logic engine embedded in a single static binary. Turn your knowledge base into queryable, verifiable reasoning.
Assumption tracking with automatic belief propagation. Retract a premise, watch the consequences unwind.
Assert facts, define rules, query goals, explain proofs, snapshot state — every tool your agent needs, exposed as MCP and CLI.
Write-ahead journal, point-in-time snapshots, automatic recovery. Zero runtime dependencies, statically linked.
# Assert facts about your system
$ zpm remember-fact "service(api, healthy)"
$ zpm remember-fact "service(db, healthy)"
$ zpm remember-fact "depends_on(api, db)"
# Define a rule
$ zpm define-rule "operational(X)" "service(X, healthy)"
# Query the knowledge base
$ zpm query-logic "operational(X)" --format json
# Explain the proof chain
$ zpm explain-why "operational(api)"
# Snapshot state before a risky change
$ zpm save-snapshot "before-deploy"Give your LLM a memory that reasons. ZPM exposes the same 22 tools over the MCP protocol and as direct CLI subcommands — use it from Claude Code, Cursor, or straight from your shell.
The reasoning engine of the AWF ecosystem. One binary, zero runtime dependencies.
curl -fsSL https://raw.githubusercontent.com/awf-project/zpm/main/scripts/install.sh | shgit clone https://github.com/awf-project/zpm.git
cd zpm
make build{
"mcpServers": {
"zpm": {
"command": "/path/to/zpm",
"args": ["serve"]
}
}
}The triggering layer of the ecosystem. A Zig-powered, push-driven scheduler — cron in reverse. Apps schedule jobs at the moment they decide, ztick fires at the requested time.
Your app sends SET over TCP at the moment it decides. ztick fires at the requested time. No polling, no fixed intervals, no drift.
Shell, direct execve, HTTP webhook, AWF workflow, AMQP publisher, Redis command. One scheduler, every backend you already run.
Native TCP protocol with TLS 1.3, plus REST/JSON HTTP API with embedded OpenAPI 3.1.1. Token AUTH, namespace-scoped authorization.
Append-only logfile with binary framing and background compression. OpenTelemetry traces and metrics over OTLP/HTTP. Zero runtime dependencies.
# Authenticate with a namespace-scoped token
$ echo "AUTH ops:s3cr3t" | nc 127.0.0.1 5678
# Define a rule: trigger an AWF workflow when the job fires
$ echo "RULE SET deploy.daily awf run release-pipeline" | nc 127.0.0.1 5678
# Schedule a one-shot job at a precise nanosecond timestamp
$ echo "SET deploy.daily $(date +%s%N -d 'tomorrow 03:00')" | nc 127.0.0.1 5678
# Inspect, list, remove
$ echo "QUERY deploy.daily" | nc 127.0.0.1 5678
$ echo "LISTRULES" | nc 127.0.0.1 5678
$ echo "REMOVE deploy.daily" | nc 127.0.0.1 5678A tight TCP protocol or a REST endpoint — pick your interface. ztick chains directly into AWF workflows, so you can trigger orchestrations on demand: no cron, no drift, no surprises.
The triggering engine of the AWF ecosystem. One Zig binary, zero runtime dependencies.
curl -fsSL https://raw.githubusercontent.com/awf-project/ztick/main/scripts/install.sh | shgit clone https://github.com/awf-project/ztick.git
cd ztick
make release
make installztick # listens on 127.0.0.1:5678
ztick -c /etc/ztick/ztick.toml # run with a TOML config
ztick dump ztick.log --follow # tail the persistence logfileA small family of projects around the CLI, ZPM, and ztick — drop AWF into your agent, your CI, or your dev container.
The Claude Code marketplace for AWF. Ship AWF to your agent in a click, with curated skills and prompts.
View on GitHubInstall AWF in any GitHub Actions workflow. One step, cached, ready for production CI.
View on GitHubA Dev Container with the full AWF stack pre-installed. Clone, open in VS Code or JetBrains IDE, start shipping.
View on GitHubExplore every repo on github.com/awf-project.
Dive into the source code and start building production-ready AI workflows with AWF.
Explore on GitHub