Frequently Asked Questions (FAQ)
Technical reference answers for common architectural, configuration, and runtime questions.
1. System Architecture
What is OpenJarvis CLI?
OpenJarvis CLI is an open-source, vendor-agnostic agentic orchestration engine and terminal client. It executes tasks by dispatching them across an asynchronous Multi-Agent System (MAS) of Conductor nodes, coordinated via message envelopes, neighbor consensus, and content-addressable artifact storage.
How does the multi-agent hierarchy operate?
The root coordinator receives the user prompt and dynamically spawns worker conductors via spawn_agent. Peer agents establish communication links via connect_agents. Agents communicate via mailboxes and must respect a bottom-up exit hierarchy: child agents must terminate via exit_agent before their parent can terminate or signal complete_task.
What is the runtime dependency footprint?
OpenJarvis CLI is pure Python (Python 3.13 / 3.14). It does not import or require PyTorch, CUDA, transformers, or GPU acceleration. Standalone binaries compiled via PyInstaller run without an external Python interpreter.
2. Providers & Models
Which LLM providers and APIs are supported?
OpenJarvis connects to any provider exposing an OpenAI-compatible /v1/chat/completions REST interface or standard provider APIs:
- Local Endpoints: Ollama, vLLM, llama.cpp, LocalAI.
- Cloud Providers: OpenAI, Anthropic, Google Gemini, Groq, OpenRouter, Mistral, DeepSeek.
Can different agents in the same session use different providers?
Yes. Each agent profile declared in .openjarvis/config.yaml can specify its own provider, model, base_url, api_key_env, temperature, and max_tokens. For instance, the root coordinator can use gpt-4o, a coding worker can run on Anthropic Claude 3.5 Sonnet, and a research worker can run against a local Ollama instance.
3. Built-In Tools & Permissions
How many built-in tools are included?
OpenJarvis registers 49 deterministic built-in tools across 9 functional categories, plus 5 multi-agent coordination meta-tools:
1. File Operations (9): read_file, write_file, list_directory, search_in_files, search_dir, search_file, find_file, file_info, delete_file
2. Git & VCS (4): git_status, git_diff, git_log, apply_patch
3. Web & Network (5): search_web, fetch_url, fetch_wikipedia, http_request, parse_openapi_spec
4. Code Execution (4): run_python, run_shell, lint_python, run_pytest
5. Code Editor & Terminal (3): str_replace_editor, execute_bash, bash
6. Data Processing (6): parse_json, jq_query, parse_csv, regex_search, regex_replace, sql_query
7. Math & Arithmetic (4): evaluate_expression, solve_equation, convert_units, prime_factorize
8. Date & Time (4): get_current_datetime, date_arithmetic, format_datetime, days_between
9. Session Memory (10): save_memory, read_memory, update_memory, delete_memory, list_memories, search_memories, and backwards-compatible note storage aliases (store_note, recall_note, list_notes, delete_note)
10. Meta-Tools (5): spawn_agent, connect_agents, report_findings, exit_agent, complete_task
How are tool permissions evaluated?
The PermissionManager enforces one of three operational modes (interactive, autonomous, allowlist). Meta-tools bypass confirmation prompts, while other tools are evaluated against blocklists, pattern rules (argument_patterns matched with fnmatch), allowlists, and optional safety classifier checks.
4. Configuration & Runtime
Where does OpenJarvis search for configuration files?
Configuration files are resolved in the following priority:
1. Explicit CLI argument: --config <path>
2. Environment variable: OJ_CONFIG
3. Workspace configuration: ./.openjarvis/config.yaml
4. User global configuration: ~/.openjarvis/config.yaml
5. System configuration: /etc/openjarvis/config.yaml
If no configuration is located, the interactive setup wizard is invoked.
How do multiline prompts work in the REPL?
In the interactive terminal, press Escape then Enter (or Alt+Enter) to insert a newline. Press Enter alone to submit the prompt.
How do non-interactive / headless runs work?
Use -p or --prompt with -y or --auto-approve (or --mode autonomous):