Orchestrate and scale AI-driven operations across your organization.
j9t is a platform for scaling AI-driven work. It transforms AI calls, Python, shell scripts, and native C++ into visual workflows that execute hundreds of tasks in parallel — enabling teams to automate complex processes and accelerate analysis, while integrating directly with their existing cloud systems and data sources.
A modern backend with multithreading, HTTP/2 multiplexing, and an adaptive rate limiter that pushes AI to the max ensures fast execution whether running locally or in the cloud. On top of this engine, j9t provides a React-based workflow editor with AI-assisted generation, structured outputs and automatic retry on failure for predictable automation, 14 cloud connectors across 5 categories, and MCP integration with external AI systems.
Workflows fire on cron schedules, file-watch events, HMAC-signed webhooks, or on demand — the same pipeline can run once from a button, every hour, or whenever a file lands in a folder.
Use cases — a single workflow can:
- Score 60 portfolio positions in seconds and roll them into one summary
- Triage a backlog of GitHub / Jira / Redmine issues into bug / feature / question buckets with automatic assignee routing
- Assess procurement requirements against a platform spec and push the verdicts back into Polarion
- Grade a classroom of quizzes and write the grades back to a Google Sheet
- Generate a diagram-rich troubleshooting PDF from schema-validated AI calls
- Describe a new workflow in plain English and let the assistant generate, validate, and auto-fix it with
branch-on-erroruntil it's valid
Why these chained pipelines beat a single prompt. Breaking a hard problem into a DAG of small AI calls gives each call a tighter scope — and higher accuracy — than one monolithic prompt. And because any step can be a compiler, make, a linter, or a test runner, deterministic tools close the loop: a failed make feeds its error back into the next AI call, which rewrites the broken file and loops until green. The same agent-loop pattern coding assistants use — encoded as a reusable, auditable pipeline your team designs once and reruns on schedule.
A dual-edition architecture lets organizations move from prototyping to secure, controlled deployment. Both editions are included in every package:
- j9t Studio — visual workflow editor with AI generation, explain, and auto-fix; AI assistant with 31 tools; workflow versioning and live debugging. Same auth posture as Engine: every browser session and MCP client requires an MCP API key.
- j9t Engine — lean production runtime. Workflow CRUD / AI assistant / AI JCWF generation are removed at compile time; everything else (run control, settings admin, log analysis, monitoring) is reachable via MCP key and role-gated (admin/operator/viewer). TLS, HMAC webhooks, rate limiting, audit log, optional gateway-header cross-check. Ready for private cloud or behind an API gateway.
Current version: 0.8.5 — working towards beta 0.95, the first major baseline subject to regression testing across all packaging targets.
| Dashboard | Workflow Editor | AI Assistant |
|---|---|---|
 |
 |
 |
JarvisAgent talks to AI providers through six interface adapters, covering every major hosted provider, every common self-hosted runtime, and the two enterprise-cloud platforms. You pick an interface in config.json by setting API to one of API1/API2/API3/API4/API5/API6.
| Adapter | Endpoint | Providers that work today |
|---|---|---|
| API1 — OpenAI Chat Completions | POST /v1/chat/completions |
OpenAI (gpt-4.1, gpt-4o, gpt-4-turbo, gpt-5, mini variants) · Google Gemini (OpenAI-compat mode) · Groq, Together AI, Fireworks, DeepInfra, Perplexity, xAI Grok, Mistral Platform, GitHub Models, OpenRouter · self-hosted: Ollama, LM Studio, llama.cpp server, vLLM, text-generation-webui |
| API2 — OpenAI Responses | POST /v1/responses |
OpenAI (Responses API — newer endpoint, used for sequential chunk throughput) |
| API3 — Gemini native | POST /v1beta/models/{model}:generateContent |
Google Gemini (native endpoint with x-goog-api-key auth) |
| API4 — Anthropic Messages | POST /v1/messages |
Anthropic Claude (Haiku / Sonnet / Opus, all generations with 200 K context) |
| API5 — AWS Bedrock | POST /model/{modelId}/invoke (SigV4-signed) |
Bedrock-hosted anthropic.claude-, meta.llama, amazon.titan-, amazon.nova- model families |
| API6 — Azure OpenAI | POST /openai/deployments/{deployment}/chat/completions?api-version={ver} |
Azure-hosted OpenAI deployments (resource-scoped URLs, api-key: header) |
Self-hosted runtimes (Ollama, LM Studio, llama.cpp, vLLM) plug into API1 — see User Manual for the config shape. Context-window handling and chunking behavior are documented in doc/architecture.md.
Workflows read from and write to external systems through a unified ICloudConnector framework. Credentials are stored encrypted, never appear in workflow files, and a single connection definition is reused across tasks.
| Category | Systems |
|---|---|
| Object storage | S3 (+ MinIO/R2/Wasabi), Azure Blob, Google Cloud Storage |
| Databases | PostgreSQL, Snowflake |
| ALM | Polarion, Jira, GitHub, Redmine |
| Collaboration | OneDrive, Google Sheets |
| Messaging | Slack, Email (SMTP/IMAP) |
Why it matters: the same automation pipeline can pull data from where it already lives, run AI analysis on it, and push results back into the systems your teams already use — no manual export/import, no separate integration scripts, no vendor lock-in.
Typical round-trip: read from cloud → fan out per item → AI processes each → write results back.
An MCP sidecar also exposes workflows to MCP clients such as Claude Code.
See doc/cloud-integration.md for the full architecture and per-connector details.
- INSTALL.md — install pre-built packages on all platforms
- DEVELOPMENT.md — build from source, dependencies, editions, running
- User Manual — MCP key enrollment, admin tasks, scripts, adhoc workflows (also available as
man jarvisagenton Linux/macOS) - doc/architecture.md — system architecture and runtime layers
- doc/JC_Workflow_Specification.md — complete JCWF format
- doc/cloud-integration.md — cloud connector framework
- doc/api-endpoints.md — REST API reference
- doc/cyber security.md — Engine security model
- integration/README.md — webhook triggers, n8n integration, HMAC signing
| Workflow | Highlights |
|---|---|
| Go-Kart Compliance Check | Polarion integration, per-item fan-out, {{template}} substitution |
| Portfolio Dividend Analysis | CSV filter, 60-position fan-out, glob aggregation |
| AI Car Maintenance Pipeline | Multi-stage pipeline with AI categorization |
| Hamburg Tourist Day Planner | Webhook trigger, n8n integration, HMAC, callback |
| Cloud round-trip demos | Polarion, PostgreSQL, S3, Email, GitHub, Jira, Slack, OneDrive, Snowflake, Google Sheets, Azure Blob, GCS, Redmine |
See INSTALL.md for full installation. Fastest path:
Prerequisites: Docker (Docker Desktop on macOS/Windows, or Docker Engine on Linux).
git clone https://github.com/beaumanvienna/JarvisAgent.git
cd JarvisAgent
./scripts/run-docker.sh # Linux / macOS
scripts\run-docker.ps1 # Windows (PowerShell)The helper script pulls ghcr.io/beaumanvienna/jarvisagent:latest and starts the container with a persistent data directory at ~/JarvisAgent.
Then open the dashboard at http://localhost:8080 (or https://localhost:8443 with TLS) and the workflow editor at /editor.
The image is published for linux/amd64 and linux/arm64 — runs natively on Intel/AMD hosts, Apple Silicon (via Docker Desktop), and ARM Linux.
Contributions are welcome. Please enable clang-format in your IDE. Coding style is Allman, and member fields of structs and classes use the m_ + PascalCase convention.
GPL-3.0 License © 2026 JC Technolabs