AI Delivery MCP

AI Delivery MCP integrates GitHub Copilot Chat in VS Code with separate community GitLab and Jira MCP servers. Copilot orchestrates delivery; the local server only prepares repository context, review content, release notes, and evidence.

Architecture

The workspace configuration in .vscode/mcp.json starts three independent stdio servers:

• aiDelivery: local TypeScript server exposing prepare_delivery, compose_jira_update, and finalize_delivery.

• gitlab: @zereight/mcp-gitlab for GitLab merge requests.

• jira: sooperset/mcp-atlassian for Jira reads, comments, and transitions.

aiDelivery does not contain GitLab/Jira REST, CLI, provider adapter, or nested MCP client code.

Related MCP server: DevOps Helper MCP

Prerequisites

• Node.js and npm

• npx

• uv with uvx

• A current VS Code release with GitHub Copilot Chat and MCP access enabled by organization policy

• GitLab token permission to create merge requests

• Jira permission to read and update the target issues

VS Code supports workspace MCP servers through .vscode/mcp.json, including stdio servers and envFile configuration. See the VS Code MCP documentation.

Setup

npm install
cp .env.example .env
npm run build

Configure .env:

GITLAB_API_URL=https://gitlab.example.com/api/v4
GITLAB_PERSONAL_ACCESS_TOKEN=replace-with-token
JIRA_URL=https://your-company.atlassian.net
JIRA_USERNAME=dev@example.com
JIRA_API_TOKEN=replace-with-token
DELIVERY_GITLAB_PROJECT_ID=group/project
DELIVERY_DEFAULT_TARGET_BRANCH=main
DELIVERY_DEFAULT_REVIEWER_IDS=1001,1002
DELIVERY_DEFAULT_LABELS=ai-delivery
JIRA_DEFAULT_DONE_TRANSITION=31

For Jira Server/Data Center, use the provider's documented JIRA_PERSONAL_TOKEN authentication instead of Jira Cloud username/API-token credentials.

.env is ignored by git. All three workspace servers load it through envFile; aiDelivery does not parse, persist, return, or log provider credentials.

Use In Copilot Chat

1. Open this repository in VS Code.

2. Run npm run build after local server code changes.

3. Open the Command Palette and run MCP: List Servers.

4. Start and trust aiDelivery, gitlab, and jira after reviewing their configurations.

5. Select the AI Delivery custom agent in Copilot Chat.

6. Run /deliver-task and provide the parent Jira issue key.

7. Review the branch range, MR payload, Jira target, release note, and evidence path.

8. To execute external writes, reply exactly CONFIRM DELIVERY <runId> using the displayed run ID.

The agent will then create one GitLab MR, compose and apply the Jira update, optionally transition Jira, and finalize local evidence. VS Code may still ask you to approve individual tool calls.

Workspace custom agents live in .github/agents, and prompt files live in .github/prompts. See the VS Code documentation for custom agents and prompt files.

Safety Model

• Jira reads and local preview generation are allowed before confirmation.

• gitlab/create_merge_request, jira/jira_update_issue, and jira/jira_transition_issue are external write tools.

• No external write is authorized until CONFIRM DELIVERY <runId> matches the current prepared run.

• If MR creation fails, Jira writes and transitions are skipped.

• Finalization records success, partial_success, or failed and recursively redacts token-shaped values.

• External result schemas reject secret-bearing keys before evidence persistence.

Local Verification

Run the complete local gate without GitLab or Jira credentials:

npm test
npm run typecheck
npm run build
npm run smoke:dry-run

The smoke command uses fixture provider results. It creates ignored artifacts under:

• artifacts/delivery-evidence/<runId>.json

• release-notes/DRY-123.md

No GitLab or Jira network calls occur during this smoke test.

Real Integration Verification

Use a test GitLab project, branch, and Jira issue first:

1. Run /deliver-task and verify no external writes occur before confirmation.

2. Confirm the exact prepared run.

3. Verify MR title, description, target branch, reviewers, and labels.

4. Verify the Jira comment and optional transition.

5. Compare the generated release note and final evidence with the real MR and Jira results.

Troubleshooting

• Run MCP: List Servers, select a server, and choose Show Output for startup or provider errors.

• Run MCP: Reset Cached Tools after changing local tool names or schemas.

• Rebuild before restarting aiDelivery: npm run build.

• Confirm GITLAB_API_URL ends with /api/v4 and DELIVERY_GITLAB_PROJECT_ID identifies the GitLab project or namespace path.

• Confirm uvx mcp-atlassian and npx -y @zereight/mcp-gitlab start successfully from a terminal if provider startup fails.

• Organization policy can disable or restrict MCP servers in Copilot Chat; contact the administrator when servers are unavailable despite valid local configuration.