Plataforma Atlassian AI Workflow
ahmet-ozel/atlassian-ai-workflow-platform
Automação de IA multi-departamento para Atlassian com workflows duráveis
Ver no GitHubMCP server que automatiza fluxos de trabalho em Jira, Confluence e Bitbucket usando agentes LLM e workflows durádouros do Temporal. Inclui gateway MCP stateless, dashboard administrativo, interface Streamlit e orquestração completa em Docker Compose, com suporte a multi-tenant e credenciais em Vault.
Instalação
npx convez-ai ahmet-ozel/atlassian-ai-workflow-platform1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Automatizar triagem e atribuição de issues Jira com análise LLM
- Gerar pull requests em Bitbucket baseado em decisões de IA
- Publicar documentação automaticamente em Confluence
- Orquestrar workflows complexos entre múltiplas aplicações Atlassian
- Governança de automação com dashboard administrativo
Veredito

Solução enterprise-grade para automação Atlassian com arquitetura sólida (Temporal, Vault, MCP stateless). Exige experiência significativa com Docker, workflows distribuídos e ecossistema Atlassian; valor real em ambientes multi-departamento com governance robusta.
README
A tradução para o português deste README ainda está na fila.
Atlassian AI Workflow Platform
Self-hosted, multi-department AI automation for Jira, Confluence and Bitbucket. Assign a Jira issue to the bot and the platform turns it into real work: it reasons over the task with an LLM, makes code changes on a runner, opens a draft Bitbucket pull request, publishes a Confluence page, and comments back on the issue - all driven by durable Temporal workflows.
It ships as a single Docker Compose stack: a stateless Atlassian MCP gateway, a webhook intake service, Temporal workers, an admin dashboard for governance, and a Streamlit chat front end for end users.
Keywords: Atlassian automation, Jira bot, Confluence automation, Bitbucket pull request automation, MCP (Model Context Protocol) server, Temporal workflows, LLM agent, self-hosted, Docker Compose.
Highlights
- Stateless, multi-tenant MCP gateway - every request carries its own Atlassian credentials in headers, so a single gateway serves many users and departments without server-side sessions.
- Secrets in Vault - department credentials and tokens are stored in HashiCorp Vault and referenced by handle; raw tokens never live in config or source.
- Admin dashboard - governance surface for departments, services, capabilities, workflows and pull-request review.
- Streamlit front end - the end-user surface for credentials, chat and task creation.
- Temporal automation - durable workflows turn a Jira event into an LLM decision, code changes on a runner, a draft pull request, a Confluence page and a Jira update.
- Env-driven host ports - every published port comes from
infra/.env, so a deployment can avoid clashes without touching Compose or source.
Quick Start
For a first-time setup we recommend bootstrap-only mode: only the
admin-dashboard and its dependencies start; every other service is activated
on demand, under operator control, from the Setup Wizard inside the dashboard.
cd platform
# infra/.env carries the bootstrap values (Postgres, Vault, host ports, LLM).
# It is git-ignored, so create it on a fresh clone before the first boot.
make boot # starts postgres + vault + admin-dashboard-{api,ui}
Then open the dashboard in your browser:
http://localhost:33000
The Setup Wizard brings up the remaining services step by step and, on the final step, asks you to add the first department.
Full walkthrough:
docs/runbooks/getting-started.md.
If make is not installed (the Windows default):
.\scripts\up.ps1 boot
or, on a POSIX shell:
./scripts/up.sh boot
Published host ports
Every service always listens on a fixed container port internally; only the
host-published port is configurable, so a deployment can avoid clashes with
other software on the same machine. All host ports are defined once in
infra/.env (*_HOST_PORT variables) - change a value there and re-run
docker compose up -d. Never edit the Compose file or application source.
| Service | Host port | Env variable |
|---|---|---|
| Admin Dashboard (UI) | 33000 | ADMIN_DASHBOARD_UI_HOST_PORT |
| Admin Dashboard API | 38082 | ADMIN_DASHBOARD_API_HOST_PORT |
| Streamlit (chat) | 38501 | STREAMLIT_HOST_PORT |
| Atlassian MCP | 38090 | ATLASSIAN_MCP_HOST_PORT |
| assistant-service | 38081 | ASSISTANT_SERVICE_HOST_PORT |
| automation-service | 38084 | AUTOMATION_SERVICE_HOST_PORT |
| task-intake-service | 38083 | TASK_INTAKE_HOST_PORT |
| Postgres | 35432 | POSTGRES_HOST_PORT |
| Redis | 36379 | REDIS_HOST_PORT |
| Vault | 38200 | VAULT_HOST_PORT |
| Temporal (gRPC) | 37233 | TEMPORAL_HOST_PORT |
| Temporal UI | 38233 | TEMPORAL_UI_HOST_PORT |
| Firecrawl | 33002 | FIRECRAWL_HOST_PORT |
| MinIO (API / console) | 39000 / 39001 | MINIO_HOST_PORT / MINIO_CONSOLE_HOST_PORT |
| Traefik (HTTP / HTTPS) | 8044 / 8444 | - (edit Compose) |
For day-to-day use you only need 33000 (dashboard) and 38501
(Streamlit chat).
Common commands
| Command | Behavior |
|---|---|
make boot | Bootstrap-only (default - postgres, vault, admin-dashboard). |
make up | Alias for make boot (backward compatibility). |
make up-all | Start every service at once (CI / debug). Not for production. |
make ps | List running services. |
make logs | Tail logs of the active services. |
make down | Stop all containers (volumes preserved). |
make restart | make down + make boot. |
make profiles | Print the profile list derived from the manifest. |
make help | Show all targets and their descriptions. |
Note on
make up-all: it starts all services at once (for CI tests and full-stack debugging). Do not use it on the first boot - launching 12 services in parallel before credentials are entered produces "started in the wrong order" errors. Follow themake boot+ Setup Wizard flow instead.
Directory layout
| Path | Contents |
|---|---|
infra/ | Docker Compose files (docker-compose.yml, docker-compose.dev.yml) and Postgres migrations. |
services/ | HTTP services - automation-service, assistant-service, admin-dashboard-api, atlassian-mcp (gateway). |
workers/ | Temporal workers - automation-worker, agent-runner-worker, execution-runner-worker. |
ui/ | Front ends - admin-dashboard (Next.js), streamlit-app (Streamlit). |
libs/ | Shared Python libraries (vault_client, audit, llm_client, etc.). |
config/ | Manifest files - services.manifest.json, departments.json and their schemas. |
prompts/ | LLM prompts - task_creation_assistant, assistant_chat, notification templates. |
docs/ | Runbooks, user guide, env reference. |
tests/ | Cross-service property/integration tests (each service also has its own tests/). |
scripts/ | up.sh, up.ps1 wrappers + maintenance scripts. |
Architecture
The stack is a set of small services around a shared MCP gateway:
- Atlassian MCP gateway (
services/atlassian_mcp_bitbucket/) - a stateless HTTP surface that exposes Jira, Confluence and Bitbucket tools. Credentials arrive per-request asX-Atlassian-*headers. - task-intake-service - receives Jira webhooks and starts workflows.
- automation-service - governance APIs (departments, credentials, PR review, branch scans) used by the admin dashboard.
- assistant-service - chat/assist APIs backed by the MCP gateway and an LLM.
- admin-dashboard (Next.js + API) - the operator/governance UI.
- streamlit-app - the end-user UI (credentials, chat, task creation).
- Temporal workers (
workers/) - durable execution for long-running automation and runner/SSH steps.
A typical automation flow:
- A Jira issue is assigned to the bot - Jira fires a webhook to
task-intake-service. - A Temporal workflow starts and asks the LLM what the task needs (code change, docs, a runner, etc.).
- The workflow drives the MCP gateway and, when needed, a remote runner over SSH to make changes.
- It opens a draft pull request on Bitbucket, publishes a Confluence page, and comments back on the Jira issue.
Built on
- Atlassian MCP gateway -
services/atlassian_mcp_bitbucket/builds thejellythomas/mcp-atlassian-with-bitbucketfork at a pinned commit. That fork extendssooperset/mcp-atlassian(the original Jira/Confluence MCP server) with Bitbucket support, so all three Atlassian products share one stateless HTTP MCP surface. - Temporal - durable workflow execution for the automation tier.
- HashiCorp Vault - secret storage for department credentials and tokens.
- Firecrawl - web content retrieval used by research/automation steps.
- Next.js (admin dashboard) and Streamlit (chat front end) for the two UIs; MinIO for artifact storage and Traefik as the optional edge router.
Next steps
- First-boot flow:
docs/runbooks/getting-started.md. - Webhook setup:
docs/runbooks/webhook-setup.md. - Department decommission:
docs/runbooks/dept-decommission.md. - Environment variables:
docs/env-reference.md. - Connect an IDE / MCP credential headers:
docs/api-contracts/mcp-credential-headers.md- Cloud vs Server/DC auth, Bearer/PAT vs Basic, and the SSRF allowlist. - End-user task-creation guide:
docs/user-guide/.
License
This repository is for private use. The bundled Atlassian MCP gateway is built from an upstream fork and retains its original upstream license.