Worka is a production-grade, deterministic AI server for orchestrating multi-agent workflows.
It acts like an AI-native nginx—serving composable DAGs, toolchains, and LLM-powered agents from your own infrastructure or via Docker.
Your multi-agent, DAG-based workflow engine powered by LLMs.
Worka transforms free-form messages into fully-automated business workflows. It empowers development teams to orchestrate complex, multi-agent workflows with confidence. Build reliable, secure, and scalable multi-agent workflows with ease. Our platform provides:
- Deterministic DAG execution: guarantee idempotent, repeatable pipelines across retries and distributed workers.
- Fine-grained context control: prune, prioritize, and shape LLM history for consistent, predictable outputs.
- Automated retry & audit: built-in error handling, low-confidence replans, and human-in-the-loop escalation.
- Comprehensive event logging: view and troubleshoot every step in the workflow’s lifecycle.
- Secure secret management & PII redaction: vault-backed runtime injection keeps your keys, tokens, and user data safe.
- Multi-agent routing & conditional logic: connect specialized agents with simple JS rules for dynamic decision-making.
- Flexible deployment: run on-premise or in the cloud via standalone API server & workers, or compile to WebAssembly for edge execution.
- TypeScript-first SDK: fully typed client with auto-generated models, code-completion, and compile-time safety.
- Multi-tenant: You control tenancy via API, mapping our internal tenants to yours.
- Secure secret management with vault-backed runtime injection and PII redaction for compliance-ready operations.
- Flexible deployment, from a standalone API server and workers to WASM-first architectures that run at the edge.
- Multi-agent routing and conditional logic powered by simple JS snippets, connecting specialized agents seamlessly.
- Built-in retry and audit, automatically handling low-confidence plans, retries, and human-in-the-loop escalation when needed.
Below is an overview of key features and their current maturity stages:
| Feature | 🗓️ Planned | 🔧 In Development | 🚀 Production | |||
|---|---|---|---|---|---|---|
| Event Logging & Inspection | 🚀 | |||||
| Retry & Audit | 🚀 | |||||
| Deterministic DAGs | 🚀 | |||||
| Semantic Keys & Nonces | 🚀 | |||||
| Multi-Agent Routing | 🚀 | |||||
| API Server & Worker Binaries | 🚀 | |||||
| Multi-Tenancy | 🚀 | |||||
| Custom Tool Integrations | 🚀 | |||||
| Core Orchestration Features | 🚀 | |||||
| Google Gemini LLM Provider | ||||||
| Google Vertex LLM Provider | ||||||
| Human-in-the-Loop Escalation | ||||||
| Fine-grained Context Control | ||||||
| TypeScript/Node.js SDK | 🚀 | |||||
| Java SDK | ||||||
| Python SDK | ||||||
| Ruby SDK | ||||||
| Rust SDK | ||||||
| ChatGPT LLM Provider | ||||||
| WASM-First Architecture | ||||||
| PII Redaction | 🔧 | |||||
| Vault Secret Management | 🔧 | |||||
| Secret Vault Management | 🔧 | |||||
| Model Context Protocol (MCP) Support | 🗓️ | |||||
| Anthropic Claude LLM Provider | 🗓 | ️ | ||||
| OLlama LLM Provider | 🗓 |
Pick the latest API server and worker binaries for your platform:
| Binary | Mac (x86_64) | Mac (ARM) | Linux (x86_64) | Linux (ARM) | Windows (x86_64) | Windows (ARM) | Docker (x86_64) | Docker (ARM) |
|---|---|---|---|---|---|---|---|---|
| API Server | api-x86_64-apple-darwin | api-aarch64-apple-darwin | api-x86_64-unknown-linux-gnu | 🚧 | api-x86_64-pc-windows-msvc.exe | api-aarch64-pc-windows-msvc.exe | 🚧 | 🚧 |
| Worker | worker-x86_64-apple-darwin | worker-aarch64-apple-darwin | worker-x86_64-unknown-linux-gnu | 🚧 | worker-x86_64-pc-windows-msvc.exe | worker-aarch64-pc-windows-msvc.exe | 🚧 | 🚧 |
- Conversation agent for initial message assignment.
- Declarative edges (
agent_edge,agent_edge_condition) define who runs next based on rules and script conditions.
- semantic_key + nonce: content-based, collision-resistant IDs.
- side_effect flag for external mutations.
- retention levels (
ephemeral,default,sticky) to manage history footprint.
max_token_budget,include_tags/exclude_tags,min_event_weight,vault_hints.- Selective history shaping ensures your prompts stay within token limits without sacrificing relevance.
- Vault table with optional conversation scoping.
vault_redactionlog for audit of Presidio-based NER redactions.$tool:vault:conversation:<key>&$tool:vault:global:<key>for secrets injection.
- logprob_observation & logprob_calibration track LLM confidence.
- Automatic DAG retries with temperature/provider fallback.
- Human escalation after configurable retry thresholds.
- Register any HTTP endpoint or inline JS as functions.
is_idempotentflag guides safe retries.- Tag-based discovery and dynamic tool injection per agent.
- ACID PostgreSQL persistence via raw SQL.
- Boa JS engine for sandboxed conditions.
- Full client- or server-side deployment in a single Wasm module.
- Install Worka on your server or embed the Wasm module in your application.
- Register your tenants, agents, rules, and tools via our APIs.
- Send your first message—Worka will spawn a workflow, generate a DAG, and execute it end-to-end.
Learn more in our documentation and see examples in the /examples folder!
This repo contains:
clients/: Official Worka SDKs and client integrations.- No source for the Worka server is included here.
- The server is distributed as a binary (e.g. Docker image).
- Clients (in this repo) are licensed under the MIT License.
- Worka Server Binary is licensed under the Worka Binary License Agreement (WBLA), which:
- Allows free use for individuals and companies with under $1M/year revenue.
- Allows redistribution with required attribution.
- Offers a $50M revenue cap if you post a tutorial or blog with attribution.
- Provides no access to or rights to the server source code.
If you redistribute Worka (e.g. bundled in apps or other Docker images), you must include a visible, do-follow link to https://worka.ai in your documentation, website, README, or UI where Worka is used.
Want to use Worka for free even if your company makes up to $50M/year?
📢 Just publish:
- A public blog post on your company or personal site describing your use of Worka.
- And/or a YouTube tutorial (≥ 5 minutes) demonstrating how Worka is used in your setup.
💡 Be sure to include a visible do-follow link to https://worka.ai in the post or video description.
