Pre-alpha software, not yet fully working. Roadmap · Issue tracker · Get in touch
How It Is Designed To Work
Steps 1-2 (pick a template, configure via wizard or REST API) work today. Steps 3-4 show the decomposition and review machinery, driven live against real models.
Choose a template
Solo Builder 3 agents 3 depts
Tech Startup 6 agents 4 depts
Engineering Squad 9 agents 3 depts
Product Studio 13 agents 6 depts
▶ Agency 13 agents 5 depts
Research Lab 8 agents 4 depts
Enterprise Org 29 agents 9 depts
Or define your own with YAML config.Built-in templates from 3 to 29 agents. Fully customisable roles, models, budgets, and tools.
POST /api/v1/setup/company
template: "agency"
company_name: "Acme AI Lab"
201 Created
agents_created: 13
departments:
- Operations (PM, Product Mgr, Scrum Master)
- Engineering (4 developers)
- Design (UX, UI)
- QA (QA Engineer, Reviewer)
- Creative (Writer, Strategist)Setup wizard or REST API. Auto-assigns LLM models, generates agent names, configures tool access.
[TASK] #12 created: "Redesign checkout flow"
[ENGINE] PM decomposed #12 into 4 leaves
[ENGINE] #12a -> UX Designer (manual)
[ENGINE] #12b -> Backend Dev (cost-aware)
[ENGINE] #12c -> Frontend Dev (smart)
[ENGINE] #12d -> QA Engineer (fastest)
[COMMS] UX Designer -> PM: "Wireframes ready"
[DONE] #12a checked by an agent that did not write it
[COMMS] Backend Dev -> QA Engineer: "Ready for review"Runtime behaviour: recursive decomposition, routing strategies, and delegation. Driven live against real models; no run has yet reached the assembly stage.
Agents active: 13 across 5 departments
Tasks completed: 47
Total cost: EUR 18.92
Coordination overhead: 14% (healthy)
Budget remaining: EUR 81.08 / 100.00
Top Performers:
Backend Dev quality: 95% tasks: 11 cost: EUR 3.20
UX Designer quality: 93% tasks: 8 cost: EUR 2.15
PM quality: 91% tasks: 9 cost: EUR 4.80Operations view (dashboard polish in active development): per-token cost tracking, coordination overhead, performance scoring, budget enforcement.
What Makes SynthOrg Different
The binding constraint is decomposition quality, not agent supply. The design puts its weight there: splitting work into genuinely independent parts, building them separately, and checking each against something that did not write it.
Orchestration Engine
A build split into a tree, driven by one engine.
Recursive decomposition into independently buildable units, four model-routing strategies, and crash recovery. Tested end to end against a deterministic harness with a scripted provider.
Sequential, parallel, Kanban, and Agile sprint workflows in the design.
Governance & Security
Every action classified, audited, and gated, by design.
A fail-closed rule engine with an LLM evaluator for the uncertain cases, four operator-set oversight modes, and budget enforcement. Wired and driven live against real models.
Autonomy is not trust: privileges are never auto-granted, and widening is a human act.
Agent Identity
Agents designed with a role, an authority, and a review.
A bound (role, model) pair per agent, authority derived from the role's place in the reporting graph, quality scoring, and hiring through offboarding with memory archival. Modelled and tested as components.
Not interchangeable workers; the reviewer cannot be the agent that wrote the work.
Memory & Knowledge
Multiple memory types. Knowledge designed to outlast any agent.
Working, episodic, semantic, procedural, and social memory with hybrid prompt and retrieval injection, built to survive offboarding and transfer to successors. Wired and durable; memory stays off until you choose an embedding model.
Organisational knowledge designed to persist across the agent lifecycle.
Architecture
The modular, protocol-driven architecture. Every component is pluggable and built as tested code.
Security-First by Design
Agents that run tools need guardrails. The pipeline and container hardening below ships today, and the runtime guardrails gate every action.
Runtime Protection
Hybrid SecOps: a fail-closed rule-engine fast-path with an LLM evaluator for uncertain cases, classified action types, and four operator-set oversight modes (full, semi, supervised, locked).
- Fail-closed rule engine + LLM fallback evaluator
- Four oversight modes, widened only by a human
- Output scanning with redaction policies
- Persistent audit trail for every decision
Hardened Infrastructure
Distroless containers, digest-pinned images, and CIS Docker Benchmark compliance. Every image is vulnerability-scanned and signed before push.
- Wolfi apko-composed distroless runtime (no shell)
- Trivy vulnerability scanning
- cosign keyless image signing
- CIS 5.3 / 5.12 / 5.25 compliance
Continuous Security
Eight security scanners run across the full development lifecycle, from first commit to production deployment. Secret detection, SAST, DAST, dependency auditing, license compliance, and supply chain scoring.
- gitleaks, CodeQL, ZAP DAST, zizmor
- pip-audit + npm audit + Socket.dev
- OSSF Scorecard supply chain maturity
- Signed commits + required reviews
Use Cases
These describe the shape of work the system is designed for.
Web application
A frontend, an API layer, and a data model built as independent leaves, then assembled bottom-up.
- Recursive task decomposition
- Isolated build workspaces
- Author barred from reviewing
Command-line tool
Argument parsing, core logic, and output formatting split into parts one agent cannot hold at once.
- Independently buildable units
- Bottom-up assembly
- Runs on your hardware
API service
Routes, request validation, and persistence become separate leaves, built together and checked before they merge.
- Concurrent leaf builds
- Reviewer separated from author
- Any provider, local included
Data pipeline
Extraction, transformation, and loading split into stages built and checked without waiting on each other.
- Dependency-aware decomposition
- Isolated workspace per leaf
- Provider-agnostic
Browser extension
Background script, content script, and settings UI as separate leaves rather than one tangle.
- Parallel leaf builds
- Never reviews its own work
- Self-hosted: no SynthOrg service in the path
Internal tool
A form, a database schema, and a small dashboard, each checked by something that did not write it.
- Recursive decomposition
- Independent review
- Models you choose
See the Dashboard
The React 19 dashboard for work, agents, approvals, and spend, with live WebSocket updates. It shows what is running, what is waiting on you, and what it costs.
License & Usage
SynthOrg uses the Business Source License 1.1: source available, production-friendly for most users, converting to Apache 2.0 automatically.
Always Free
- Personal learning and experimentation
- Academic and research use
- Development, testing, and evaluation
- Contributing to SynthOrg itself
Free in Production
- Organizations with <500 employees and contractors (including affiliates)
- Non-competing use: internal tools, automation, or products where SynthOrg is not the core value
- No separate license needed; see LICENSE for full terms
Commercial License
- Organizations with 500+ employees
- Offering SynthOrg as a hosted service
- Reselling or embedding as core product
- May be granted at no cost; just ask
Every version converts to Apache 2.0 automatically. We welcome feedback on the license: start a discussion or open an issue.
Read the full licensing rationale →Keep the build on your own machine.
Get StartedGet in Touch
Questions about licensing, commercial use, or the framework? We'd love to hear from you.
Message sent
Thanks for reaching out! We'll get back to you soon.