Agentic Jira automation platform combines MCP and Temporal for durable workflows

Ahmet Özel describes a recurring failure mode in AI automation: agent loops that work fine for short tasks but collapse when a workflow spans minutes or hours, depends on intermittently failing external systems, or needs to survive a process restart. His solution is an open-source platform that separates three responsibilities — the agent handles intent, MCP owns the tool boundary, and Temporal owns reliability.

Temporal wraps each LLM call and each tool call as a checkpointed activity; if a worker dies, the workflow resumes from the last completed step on another worker.

The MCP integration layer exposes Jira and Confluence as typed tools (create issue, transition status, search, comment, fetch a Confluence page), decoupling tool definitions from agent logic so either side can change independently. Temporal wraps each LLM call and each tool call as a checkpointed activity; if a worker dies, the workflow resumes from the last completed step on another worker. Long-running human-approval steps — waiting for a reviewer before transitioning a ticket — become a first-class workflow primitive rather than a workaround. The acknowledged tradeoff is infrastructure overhead: Temporal is an additional service, and workflow code must be strictly deterministic, with all non-deterministic operations (network calls, timestamps, random values) isolated inside activities rather than the workflow body.

The full stack runs in a single Docker Compose setup and includes a webhook gateway that converts Jira events into workflow triggers, a Streamlit UI for running and inspecting workflows, and config-driven multi-provider LLM support covering OpenAI, Anthropic, Gemini, and self-hosted vLLM. Özel notes the tool layer is designed to extend beyond Atlassian to other platforms, and invites discussion on how others handle long-running agent workflows and multi-system MCP tool structuring.

Key facts

1. 01Built by Ahmet Özel; open-source repo at github.com/ahmet-ozel/atlassian-ai-workflow-platform
2. 02Uses MCP to expose Jira and Confluence as typed tools (create issue, transition status, search, comment, fetch page), decoupling tools from agent logic
3. 03Temporal checkpoints every LLM call and tool call as an activity, so only the failed step is retried on crash — not the full reasoning chain
4. 04Non-deterministic operations (network calls, timestamps, random values) must live in Temporal activities, not workflow body code
5. 05Supports OpenAI, Anthropic, Gemini, and self-hosted vLLM via config-driven provider selection
6. 06Includes a webhook gateway for Jira event triggers and a Streamlit UI for workflow inspection
7. 07Entire stack runs in a single Docker Compose setup

Topics