Configuration
Configuration (environment variables)
Authoritative reference: agentic-orchestration-tool/.env.example — every variable is documented inline with defaults and behavior.
Load order: python-dotenv loads .env from the tool directory when you start main.py (see main.py).
Categories (summary)
| Category examples | Notes | |
|---|---|---|
| OpenAI | OPENAI_API_KEY, OPENAI_MODEL_NAME, OPENAI_BASE_URL (compatible servers) |
|
| Anthropic | ANTHROPIC_API_KEY, ANTHROPIC_BASE_URL |
|
| Hugging Face | HF_TOKEN, HUGGINGFACE_API_BASE |
|
| Ollama | OLLAMA_HOST, router model envs |
|
| Dynamic planner | AGENTIC_PLANNER_MODEL, AGENTIC_PLANNER_USE_LITELLM, AGENTIC_PLANNER_MAX_STEPS, JSON mode, repair retry, 429 retries, context window truncation |
|
| Sessions | AGENTIC_ORCHESTRATOR_* |
|
| VRAM / hardware | AGENTIC_ASSUME_VRAM_GB, AGENTIC_MAX_VRAM_FRACTION, AGENTIC_MAX_VRAM_GB, disable filters |
|
| MCP | AGENTIC_EXTRA_MCP_PROVIDERS_PATH, HOME_ASSISTANT_*, BRAVE_SEARCH_*, TAVILY_API_KEY, EXA_API_KEY, AGENTIC_MCP_FETCH_ENABLED, AGENTIC_MCP_MEMORY_MCP_ENABLED, FILESYSTEM_MCP_ALLOWED_DIRECTORY, goal-match toggles |
|
| Agent skills | AGENTIC_AGENT_SKILLS_CATALOG, AGENTIC_EXTRA_AGENT_SKILLS_PATH, AGENTIC_SKILLS_MAX_CHARS_PER_TASK, AGENTIC_DISABLE_SKILL_GOAL_MATCH, AGENTIC_STRICT_SKILL_IDS |
|
| Agent harness | AGENTIC_HARNESS_TIER, AGENTIC_HARNESS_EVAL, AGENTIC_HARNESS_EVAL_MODEL, AGENTIC_HARNESS_SKIP_SELFCONTAINED_INIT, AGENTIC_HARNESS_FEED_PLANNER, AGENTIC_HARNESS_RECORD_STATS — see Agent harness roadmap |
|
| Execution backend | AGENTIC_EXECUTION_BACKEND, AGENTIC_SUBPROCESS_WORKERS, AGENTIC_RUN_STORE_PATH |
See Dual execution framework, Kubernetes execution upgrade |
| Progress / step context | AGENTIC_PROGRESS, AGENTIC_STEP_CONTEXT_* |
|
| Learning & KB | AGENTIC_LEARNING*, AGENTIC_KB* (attachment fingerprints: attachment_fingerprint; legacy mcp_fingerprint alias) |
|
| Answer cache | AGENTIC_ANSWER_CACHE |
|
| Iterative mode | AGENTIC_DYNAMIC_ITERATIVE_*, controller-related vars |
|
| Iterative stdout behavior | AGENTIC_DYNAMIC_ITER_STREAM_STEPS |
|
| Attachments / uploads | AGENTIC_ATTACHMENTS_ALLOW_ABSOLUTE (advanced escape hatch; defaults keep paths under tool upload directory) |
|
| Extra catalogs | AGENTIC_EXTRA_AGENT_PROVIDERS_PATH, AGENTIC_EXTRA_MCP_PROVIDERS_PATH, AGENTIC_EXTRA_AGENT_SKILLS_PATH |
|
| Artifacts | AGENTIC_VERIFY, output dirs |
Execution backend
| Variable | Default | Meaning |
|---|---|---|
AGENTIC_EXECUTION_BACKEND |
inprocess |
inprocess (CrewAI in-process), subprocess, or kubernetes (stub). Aliases: crewai, k8s. |
AGENTIC_SUBPROCESS_WORKERS |
0 |
When 1 and backend is subprocess, run each step via python main.py --execute-step. Otherwise subprocess backend falls back to in-process CrewAI. |
AGENTIC_RUN_STORE_PATH |
(unset) | Mounted directory for distributed step handoffs: {path}/{run_id}/{step_id}/result.json. Used by subprocess workers today; same path on a PVC for K8s Jobs. When unset, each run uses a temp directory. |
AGENTIC_K8S_ALLOW_STDIO_MCPS |
0 |
K8s mode only: when 1, allow stdio MCP ids if sidecars exist (K4). K3 MVP keeps 0. See Kubernetes execution upgrade. |
See Dual execution framework for architecture and phased rollout. See Kubernetes execution upgrade for PVC layout on cluster.
Platform agent harness
| Variable | Default | Meaning |
|---|---|---|
AGENTIC_HARNESS_TIER |
static |
Default tier when --harness-tier omitted: static, connectivity, smoke, capability. |
AGENTIC_HARNESS_EVAL |
1 |
When 0, skip LLM rubric on L3 capability tier. |
AGENTIC_HARNESS_EVAL_MODEL |
(planner default) | Model for capability-tier eval. |
AGENTIC_HARNESS_SKIP_SELFCONTAINED_INIT |
0 |
When 1, L1 connectivity skips initialize() on Ollama selfcontained entries (faster CI). |
AGENTIC_HARNESS_FEED_PLANNER |
1 |
Inject recent harness failures into dynamic planner context. |
AGENTIC_HARNESS_RECORD_STATS |
1 |
Record harness pass/fail in __orchestrator_learning__/stats.json. |
CLI: --harness-agent, --harness-batch, --harness-filter, --harness-json. Helpers: scripts/run-agent-harness.ps1, scripts/harness-report.py.
Notable runtime toggles
AGENTIC_ANSWER_CACHE=1(default): same-goal replay in the same orchestrator session, with explicit “reply no to re-run” flow.AGENTIC_DYNAMIC_ITER_STREAM_STEPS=1: emit each iterative round output to stdout instead of only the final synthesis.AGENTIC_OLLAMA_PULL_PROGRESS_STDERR=1(default): keep normalized Ollama pull progress lines visible on stderr for web activity/progress UI.
Web server (agentic-orchestration-web/.env)
| Variable | Role |
|---|---|
AGENTIC_TOOL_ROOT |
Path to folder containing main.py |
AGENTIC_PYTHON |
Python executable (defaults to tool .venv when present) |
AGENTIC_WEB_HOST |
Bind address (0.0.0.0 for LAN) |
AGENTIC_WEB_PORT |
Default 3847 |
See agentic-orchestration-web/README.md.
Security
Never commit .env or tokens. .env files are gitignored by convention.
Related
- MCP providers — required env per integration; awesome-mcp-servers cross-reference
- Agent harness roadmap — tiers, profiles, CI
- Agent skills roadmap — procedural skill catalog env vars and attachment semantics
- Agent skills — shipped skill inventory
- CLI reference
- Root
README.md— summary table