Environment Variables Reference

This is the complete reference for all Manasvi configuration settings. All settings go in your .env file.

Core security keys

These settings control how Manasvi services authenticate with each other internally. They are required.

Variable	Description	Example
`INTERNAL_AUTH_KEY_ID`	ID of the key used to sign internal tokens	`local-k1`
`INTERNAL_AUTH_SIGNING_SECRET`	Secret used to sign internal tokens (min 32 chars in production)	`my-local-secret-at-least-32-chars`
`INTERNAL_AUTH_VERIFICATION_KEYS`	All valid keys, as `keyId:secret` pairs separated by commas	`local-k1:my-secret`
`INTERNAL_AUTH_ISSUER`	Token issuer identifier	`manasvi.internal.auth`
`INTERNAL_AUTH_AUDIENCE`	Token audience identifier	`manasvi.internal.services`

Approval keys

Used to sign and verify approval artifacts.

Variable	Description	Example
`APPROVAL_SIGNING_KEY_ID`	Active key ID for signing approvals	`approval-k1`
`APPROVAL_SIGNING_KEYS`	Key pairs `id:secret`, comma-separated	`approval-k1:approval-secret`
`APPROVAL_VERIFICATION_KEYS`	Same format, all valid approval keys	`approval-k1:approval-secret`

Model provider

Variable	Description	Default
`MANASVI_MODEL_PROVIDER`	Default provider (`deepseek`, `openai`, `ollama`, `claude`, `mock`)	`deepseek`
`MANASVI_MODEL`	Default model id	`deepseek-v4-flash`
`MODEL_ADAPTER_MODE`	`deepseek`, `openai`, `ollama`, `claude`, `mock`, or `auto`	`deepseek`
`PLANNER_MODEL`	Model name to use	`deepseek-v4-flash`
`DEEPSEEK_API_KEY`	DeepSeek API key	—
`DEEPSEEK_BASE_URL`	DeepSeek API base URL	`https://api.deepseek.com`
`DEEPSEEK_TIMEOUT_MS`	DeepSeek request timeout (ms)	`60000`
`OPENAI_API_KEY`	OpenAI API key	—
`OPENAI_BASE_URL`	OpenAI-compatible API base URL	`https://api.openai.com/v1`
`ANTHROPIC_API_KEY`	Anthropic API key	—
`ANTHROPIC_BASE_URL`	Anthropic API base URL	`https://api.anthropic.com`
`OLLAMA_BASE_URL`	Ollama API base URL	`http://localhost:11434/v1`
`MODEL_ADAPTER_TIMEOUT_MS`	Model request timeout (ms)	`60000`
`MODEL_ADAPTER_MAX_CONTEXT_CHUNKS`	Max context chunks to include	`24`

Service URLs

How services find each other. Defaults work for local development.

Variable	Default
`EVENT_BUS_TARGET_URLS`	`http://localhost:4102/internal/events`
`POLICY_SERVICE_BASE_URL`	`http://localhost:4103`
`APPROVAL_SERVICE_BASE_URL`	`http://localhost:4108`
`EXECUTION_MANAGER_BASE_URL`	`http://localhost:4104`
`MEMORY_SERVICE_BASE_URL`	`http://localhost:4105`
`ORCHESTRATOR_BASE_URL`	`http://localhost:4102`
`NODE_MANAGER_BASE_URL`	`http://localhost:4106`

Service ports

Override the default port for any service.

Variable	Default	Service
`SERVICE_PORT`	Varies	Set per service
`SERVICE_HOST`	`0.0.0.0`	Bind address

Channel integrations

Variable	Description
`TELEGRAM_BOT_TOKEN`	Token from BotFather
`TELEGRAM_WEBHOOK_SECRET`	Random secret for webhook verification

Slack

Variable	Description
`SLACK_BOT_TOKEN`	Bot user OAuth token (`xoxb-...`)
`SLACK_SIGNING_SECRET`	Signing secret from Slack app settings

Google integration foundation

Variable	Description	Default
`GOOGLE_OAUTH_CLIENT_ID`	Google OAuth client ID	—
`GOOGLE_OAUTH_CLIENT_SECRET`	Google OAuth client secret	—
`GOOGLE_OAUTH_REDIRECT_URI`	OAuth callback URL handled by api-gateway	`http://127.0.0.1:4100/integrations/oauth/google/callback`
`INTEGRATIONS_STATE_FILE`	File path for OAuth state records	`.manasvi/integrations/oauth-states.json`
`INTEGRATIONS_ACCOUNTS_FILE`	File path for integration account metadata	`.manasvi/integrations/accounts.json`
`INTEGRATIONS_TOKENS_FILE`	File path for encrypted token vault	`.manasvi/integrations/tokens.enc.json`
`INTEGRATIONS_TOKEN_ENCRYPTION_SECRET`	32+ char secret used to encrypt token vault records	falls back to `MEMORY_ENCRYPTION_KEY`

Execution and sandbox

Variable	Description	Default
`EXECUTION_INTENT_TTL_SECONDS`	How long before an intent expires	`900`
`APPROVED_ARTIFACT_TTL_SECONDS`	How long an approval is valid	`900`
`APPROVAL_REQUEST_TTL_SECONDS`	How long before a pending approval expires	`3600`
`SANDBOX_ROOT_DIR`	Filesystem root for sandboxed executions	`/tmp/manasvi-runs`
`SANDBOX_MAX_OUTPUT_BYTES`	Maximum tool output size	`65536`
`EXECUTION_SHELL_ALLOWED_COMMANDS`	Server-side allowlist for the shell-command tool (comma-separated). Callers cannot supply this. Empty = shell tool denied (fail-closed).	(empty)
`MCP_SERVERS_JSON`	MCP servers to host and expose as governed tools. See MCP Tools. MCP tools are approval-gated unless declared read-only.	(empty)
`EXECUTION_REPLAY_STORE_DIR`	Directory for durable consumed artifacts, nonces, and idempotency keys in execution-manager. Empty keeps in-memory behavior.	(empty)
`EXECUTION_REPLAY_RETENTION_SECONDS`	Retention window for execution-manager replay records. Keep this longer than approval/intent TTLs.	`3600`
`EXECUTION_SANDBOX_CONTAINERIZED`	Set to `true` only when execution-manager is actually running inside an OS-isolated sandbox backend. Required for tools that declare native isolation needs.	`false`

Durability and audit

Variable	Description	Default
`MEMORY_PERSISTENCE_FILE`	Optional JSONL persistence file for the memory plane. Sensitive records remain encrypted at rest.	(empty)
`NODE_MANAGER_REPLAY_STORE_DIR`	Directory for durable node-manager dispatch nonce records. Empty keeps in-memory behavior.	(empty)
`NODE_AGENT_REPLAY_STORE_DIR`	Directory for durable node-agent dispatch/artifact replay records. Empty keeps in-memory behavior.	(empty)
`APPROVAL_STATE_FILE`	Optional JSON snapshot store for approval requests, records, artifacts, and resume callback state. Required for approval queue durability across restarts.	(empty)
`APPROVAL_RESUME_CALLBACK_ENABLED`	Enables approval-service to call orchestrator `/agent-runtime/resume` after approve/reject decisions.	`true`
`APPROVAL_RESUME_CALLBACK_INTERVAL_MS`	Retry interval for pending or failed approval resume callbacks.	`5000`
`ORCHESTRATOR_AGENT_RUN_STORE_FILE`	Optional JSONL store for paused agent runs awaiting approval. Required for approval resume across orchestrator restarts.	(empty)
`AUDIT_SERVICE_BASE_URL`	Base URL used by services that forward audit events.	`http://localhost:4107`
`POLICY_AUDIT_FORWARDING_ENABLED`	Set to `false` to stop policy-service from forwarding decisions to audit-service. Keep enabled for governance evidence.	`true`

Logging

Variable	Description	Default
`LOG_LEVEL`	`debug`, `info`, `warn`, `error`	`info`
`HUMAN_LOGS`	Set to `true` for human-readable logs	`false`
`SERVICE_VERSION`	Version string in logs	`0.2.0`

Agent loop behavior

Variable	Description	Default
`AGENT_LOOP_MAX_ITERATIONS`	Max planning iterations per request	`6`
`AGENT_LOOP_MAX_CONSECUTIVE_FAILURES`	Max consecutive errors before halt	`2`
`AGENT_LOOP_MAX_SUBAGENT_DEPTH`	Maximum child-agent depth for `tool.subagents`	`2`
`AGENT_LOOP_MAX_SUBAGENTS_PER_TURN`	Maximum subagents a parent turn may start	`3`
`AGENT_LOOP_MAX_SUBAGENT_ITERATIONS`	Maximum iterations for each child turn, capped at `4`	`2`
`SESSION_CONTEXT_TOKEN_BUDGET`	Context window limit (tokens)	`2048`
`SESSION_RECENT_MESSAGE_LIMIT`	Recent messages to include in context	`20`

Tips

Never commit .env to version control. It contains secrets.
For local development, simple consistent strings work fine for keys.
In production, use random 32+ character secrets.
Restart services after changing .env — settings are read at startup.

Core security keys​

Approval keys​

Model provider​

Service URLs​

Service ports​

Channel integrations​

Telegram​

Slack​

Google integration foundation​

Execution and sandbox​

Durability and audit​

Logging​

Agent loop behavior​

Tips​