Developer features¶

If you are a developer, Jarvis becomes your AI engineering hub: cross-IDE coding assistant, code review, repository automation, GitHub integration.

What you can do¶

• 💻 Coding assistant integrated in VS Code, JetBrains, terminal
• 🔍 Code review with Jarvis as "second pair of eyes"
• 🧪 Automated test generation with TDD-first methodology
• 🚀 Voice CI/CD trigger ("Hey Jarvis, deploy staging")
• 📋 PR automation (changelog, description, issue links)
• 🐛 Bug investigation with semantic grep over the codebase
• 📦 Dependency review and security audit

Stack: open-source coding assistants¶

Recommendation: Cline (VS Code) + Aider (Git-focused terminal) as the main stack. Goose as fallback for tasks needing system-command execution.

Developer Agent architecture¶

flowchart LR
    DEV[Developer] --> IDE[VS Code · JetBrains · Terminal]
    IDE -->|MCP| JAGENT[Jarvis Dev Agent]
    JAGENT --> CTX[Context retriever]
    CTX --> REPO[Repo files]
    CTX --> RAG[RAG on docs]
    CTX --> GIT[Git history]
    CTX --> ISSUE[GitHub issues/PR]
    JAGENT --> LLM[LLM router]
    LLM --> LOCAL[Qwen2.5-Coder · DeepSeek-Coder]
    LLM --> CLOUD[Claude Sonnet · GPT-4]
    JAGENT --> ACT[Action layer]
    ACT --> EDIT[Edit file]
    ACT --> RUN[Run command]
    ACT --> TEST[Run tests]
    ACT --> PR[Open PR]

Configuration¶

# Coding model — local via Ollama
JARVIS_MODEL_CODE=ollama/qwen2.5-coder:14b

# Cloud fallback for hard tasks
ANTHROPIC_API_KEY=sk-ant-...
JARVIS_MODEL_CODE_CLOUD=anthropic/claude-sonnet-4-6

# GitHub integration
GITHUB_PAT=ghp_...
GITHUB_DEFAULT_REPO=fedcal/open-jarvis

Custom MCP server¶

Jarvis exposes its own MCP server (server/orchestration/mcp_server.py) providing Cline, Continue, Goose, Claude Code with these tools:

• repo.search(query) — semantic search across the codebase
• repo.diff() — current diff
• tests.run(suite) — run specific tests
• gh.create_pr(branch, title, body) — open PR
• gh.list_issues(state) — repo issues
• memory.recall(query) — recall personal memory
• docs.lookup(library, topic) — Context7 lookup

Configuration in .cline/mcp_settings.json (example):

{
  "mcpServers": {
    "jarvis": {
      "url": "http://localhost:8080/mcp",
      "headers": { "Authorization": "Bearer ${JARVIS_TOKEN}" }
    }
  }
}

Usage examples¶

Code review from chat¶

You: Take a look at PR #42 and tell me if you see issues
Jarvis: I read the 7 modified files. Three findings:
        1. validate_input in utils.py is missing a null check (line 23)
        2. The new migration creates an index without CONCURRENTLY
        3. Integration test test_user_create does not isolate the DB
        Want me to comment on the PR?

TDD test generation¶

You: Create tests for calculate_tax in tax_utils.py
Jarvis: I generated 6 tests:
        - happy path
        - zero rate
        - negative rate (must raise ValueError)
        - zero amount
        - negative amount
        - decimal precision (snapshot)
        Saved in tests/test_tax_utils.py. Run them?

Bug investigation¶

You: Since yesterday the deploy fails with "connection refused" to Redis
Jarvis: I checked:
        - git log last 24h: 3 commits, none touch Redis
        - docker compose ps: redis container is 'unhealthy'
        - redis log: "MISCONF Redis is configured to save RDB snapshots..."
        Cause: disk full. Free space or expand the volume.

Daily standup¶

You: What did I do yesterday on project X?
Jarvis: 4 commits, 2 merged PRs, 1 closed issue.
        - feat(memory): added TTL to short-term cache
        - fix(auth): handle OAuth refresh race condition
        - PR #84: voice-agent wake word retraining
        - Issue #91 closed with user confirmation

Recommended workflow¶

1. Open VS Code with Cline configured
2. Parallel terminal with Aider for Git-focused sessions
3. Jarvis Dev Agent in the background for:
   - Automatic daily standup
   - Cross-IDE persistent memory
   - Notifications on CI failures
   - Architectural decision recording

Code privacy¶

• ✅ 100% local: use Ollama + Qwen2.5-Coder. Not a single byte of code leaves
• ⚠️ Hybrid: cloud LLM only for selected tasks (architectural review, design doc generation)
• ❌ Never upload corporate repo code to cloud LLMs without explicit authorisation

Roadmap¶