Environment Variables

All environment variables are read at startup. Set them in your shell, .env file, or MCP client configuration.

Skill Gating

These variables unlock optional skill domains that are off by default.

`ENABLE_PHYSICS_SKILLS`

Default	Values
`false`	`true` / `false`

Enables the qm-* (Quantum Mechanics) and gr-* (General Relativity) skill prefixes. Without this flag, any attempt to invoke physics skills returns an authorization error.

ENABLE_PHYSICS_SKILLS=true node dist/index.js

`ALLOW_GOVERNANCE_SKILLS`

Default	Values
`false`	`true` / `false`

Enables the gov-* governance and compliance skill prefix. The govern workflow and policy-govern tool require this flag. Designed for regulated environments where governance tooling must be explicitly opted-in.

ALLOW_GOVERNANCE_SKILLS=true node dist/index.js

`DISABLE_ADAPTIVE_ROUTING`

Default	Values
`false`	`true` / `false`

Disables the adapt-* bio-inspired routing skill prefix and hides the routing-adapt MCP tool from the public surface. By default routing-adapt is visible and adaptive skills are enabled — set this flag to opt-out. Hebbian reinforcement, ACO pheromone routing, Physarum pruning, and simulated annealing workflows remain available unless this flag is set.

# Opt-out: hide routing-adapt and block adapt-* skill execution
DISABLE_ADAPTIVE_ROUTING=true node dist/index.js

`ALLOW_INTENSIVE_SKILLS`

Default	Values
`false`	`true` / `false`

Enables long-horizon, compute-intensive skill variants that may exceed standard token budgets or incur higher cost. Consult skill documentation before enabling in production.

Tool Surface Control

`HIDDEN_TOOLS`

Default	Values
`""`	Comma-separated tool names

Removes specific tools from the tools/list response. Useful for restricting the tool surface in shared or embedded configurations.

# Hide physics and governance tools from the tool surface
HIDDEN_TOOLS="physics-analysis,enterprise,govern" node dist/index.js

Runtime & Debug

`NODE_ENV`

Default	Values
`development`	`development` / `production` / `test`

Controls environment-specific behaviour including observability verbosity and validation strictness. Set to production for deployed servers.

`VITEST`

Default	Values
`false`	`true` / `false`

Automatically set by the test runner. Enables test-safe model availability stubs and skips live network checks.

`VALIDATION_MODE`

Default	Values
`strict`	`strict` / `advisory`

In advisory mode, input validation failures are logged as warnings rather than rejected errors. Useful for local development and prompt iteration.

Serena Integration

See Serena Integration for the full backend description. By default the server runs in advisory mode and emits structured hints that the host model executes via its own Serena connection. The variables below opt into the child-spawn mode instead, where this server starts Serena as a stdio subprocess and resolves queries directly.

`MCP_SERENA_COMMAND`

Default	Values
unset	shell command, e.g. `uvx`

When set, the server lazily spawns Serena as a child MCP server over stdio and proxies queries to it. When unset, the server falls back to advisory mode and emits suggested mcp__serena__* calls for the host model to execute.

`MCP_SERENA_ARGS`

Default	Values
unset	space-separated args

Passed verbatim to MCP_SERENA_COMMAND. Recommended:

MCP_SERENA_ARGS="--from git+https://github.com/oraios/serena serena start-mcp-server --project /absolute/path/to/your/project"

Pin --project <path> explicitly: Serena keeps a global registry and otherwise refuses memory operations on a fresh cwd.

`MCP_SERENA_CWD`

Default	Values
parent cwd	absolute path

Working directory for the spawned Serena child.

`MCP_LOCAL_MEMORY`

Default	Values
`false`	`true` / `false`

Opt-in switch for the legacy per-tool-call TOON memory artifacts under .mcp-ai-agent-guidelines/memory/. Off by default — Serena memory (advertised via the 🧭 enrichment footer) is the recommended cross-session channel.

Model Router

Model router variables are managed via orchestration.toml and the orchestration-model-discover MCP tool. Avoid hardcoding model display names — use role classes (free, cheap, strong, reviewer) in code and let the router resolve actual model IDs at runtime.

See Model Routing for the full role hierarchy and discovery workflow.