| Claude | Gemini | OpenAI | Llama |
|---|---|---|---|
- 🤖 Atlassian Agent is an LLM-powered agent built using the LangGraph ReAct Agent workflow and MCP tools.
- 🌐 Protocol Support: Compatible with A2A protocol for integration with external user clients.
- 🛡️ Secure by Design: Enforces Atlassian API token-based RBAC and supports external authentication for strong access control.
- 🔌 Integrated Communication: Uses langchain-mcp-adapters to connect with the Atlassian MCP server within the LangGraph ReAct Agent workflow.
- 🏭 First-Party MCP Server: The MCP server is generated by our first-party openapi-mcp-codegen utility, ensuring version/API compatibility and software supply chain integrity.
- Ensure your
.envfile is set up as described in the cnoe-agent-utils usage guide based on your LLM Provider. - Refer to .env.example as an example.
Example .env file:
LLM_PROVIDER=
AGENT_NAME=atlassian
ATLASSIAN_TOKEN=
ATLASSIAN_EMAIL=
ATLASSIAN_API_URL=
ATLASSIAN_VERIFY_SSL=
########### LLM Configuration ###########
# Refer to: https://github.com/cnoe-io/cnoe-agent-utils#-usageUse the following link to get your own Atlassian API Token:
https://id.atlassian.com/manage-profile/security/api-tokens
Run the agent in a Docker container using your .env file:
docker run -p 0.0.0.0:8000:8000 -it\
-v "$(pwd)/.env:/app/.env"\
ghcr.io/cnoe-io/agent-atlassian:a2a-stableUse the agent-chat-cli to interact with the agent:
uvx https://github.com/cnoe-io/agent-chat-cli.git a2aflowchart TD
subgraph Client Layer
A[User Client A2A]
end
subgraph Agent Transport Layer
B[Google A2A]
end
subgraph Agent Graph Layer
C[LangGraph ReAct Agent]
end
subgraph Tools/MCP Layer
D[LangGraph MCP Adapter]
E[Atlassian MCP Server]
F[Atlassian API Server]
end
A --> B --> C
C --> D
D -.-> C
D --> E --> F --> E
- 🤖 LangGraph + LangChain MCP Adapter for agent orchestration
- 🧠 Azure OpenAI GPT-4o as the LLM backend
- 🔗 Connects to Atlassian via a dedicated Atlassian MCP agent
If you don't have an existing Atlassian server, you can quickly spin one up using Minikube:
- Start Minikube:
minikube start- Install Atlassian in the
atlassiannamespace:
kubectl create namespace atlassian
kubectl apply -n atlassian -f https://raw.githubusercontent.com/argoproj/argo-cd/stable/manifests/install.yaml- Expose the Atlassian API server:
kubectl port-forward svc/atlassian-server -n atlassian 8080:443The API will be available at https://localhost:8080.
- Get the Atlassian admin password:
kubectl -n atlassian get secret atlassian-initial-admin-secret -o jsonpath="{.data.password}" | base64 -d && echo- (Optional) Install Atlassian CLI:
brew install atlassian
# or see https://argo-cd.readthedocs.io/en/stable/cli_installation/For more details, see the official getting started guide.
To interact with the agent in A2A mode:
make run-a2a-clientSample Streaming Output
When running in A2A mode, you'll see streaming responses like:
============================================================
RUNNING STREAMING TEST
============================================================
--- Single Turn Streaming Request ---
--- Streaming Chunk ---
The current version of Atlassian is **v2.13.3+a25c8a0**. Here are some additional details:
- **Build Date:** 2025-01-03
- **Git Commit:** a25c8a0eef7830be0c2c9074c92dbea8ff23a962
- **Git Tree State:** clean
- **Go Version:** go1.23.1
- **Compiler:** gc
- **Platform:** linux/amd64
- **Kustomize Version:** v5.4.3
- **Helm Version:** v3.15.4+gfa9efb0
- **Kubectl Version:** v0.31.0
- **Jsonnet Version:** v0.20.0
- 🛠️ Uses
create_react_agentfor tool-calling - 🔌 Tools loaded from the Atlassian MCP server (submodule)
- ⚡ MCP server launched via
uv runwithstdiotransport - 🕸️ Single-node LangGraph for inference and action routing
agent_atlassian/
│
├── agent.py # LLM + MCP client orchestration
├── langgraph.py # LangGraph graph definition
├── __main__.py # CLI entrypoint
├── state.py # Pydantic state models
└── atlassian_mcp/ # Git submodule: Atlassian MCP server
This project uses a first-party MCP module generated from the Atlassian OpenAPI specification using our openapi-mcp-codegen utility. The generated MCP server is included as a git submodule in atlassian_mcp/.
All Atlassian-related LangChain tools are defined by this MCP server implementation, ensuring up-to-date API compatibility and supply chain integrity.
The agent uses MultiServerMCPClient to communicate with MCP-compliant services.
Example (stdio transport):
async with MultiServerMCPClient(
{
"atlassian": {
"command": "uv",
"args": ["run", "/abs/path/to/atlassian_mcp/server.py"],
"env": {
"ATLASSIAN_TOKEN": atlassian_token,
"ATLASSIAN_API_URL": atlassian_api_url,
"ATLASSIAN_VERIFY_SSL": "false"
},
"transport": "stdio",
}
}
) as client:
agent = create_react_agent(model, client.get_tools())Example (SSE transport):
async with MultiServerMCPClient(
{
"atlassian": {
"transport": "sse",
"url": "http://localhost:8000"
}
}
) as client:
...This evaluation uses agentevals to perform strict trajectory match evaluation of the agent's behavior. To run the evaluation suite:
make evalsThis will:
- Set up and activate the Python virtual environment
- Install evaluation dependencies (
agentevals,tabulate,pytest) - Run strict trajectory matching tests against the agent
=======================================
Setting up the Virtual Environment
=======================================
Virtual environment already exists.
=======================================
Activating virtual environment
=======================================
To activate venv manually, run: source .venv/bin/activate
. .venv/bin/activate
Running Agent Strict Trajectory Matching evals...
Installing agentevals with Poetry...
. .venv/bin/activate && uv add agentevals tabulate pytest
...
set -a && . .env && set +a && uv run evals/strict_match/test_strict_match.py
...
Test ID: atlassian_agent_1
Prompt: show atlassian version
Reference Trajectories: [['__start__', 'agent_atlassian']]
Note: Shows the version of the Atlassian Server Version.
...
Results:
{'score': True}
...
Latest Strict Match Eval Results
Apache 2.0 (see LICENSE)
See MAINTAINERS.md
- Contributions welcome via PR or issue!
- LangGraph and LangChain for agent orchestration frameworks.
- langchain-mcp-adapters for MCP integration.
- AGNTCY Agent Gateway Protocol(AGP)
- AGNTCY Workflow Server Manager (WFSM) for deployment and orchestration.
- Model Context Protocol (MCP) for the protocol specification.
- Google A2A
- The open source community for ongoing support and contributions.