Environment variables

Two kinds of thing live here. Credentials, which live nowhere else, and overrides, which beat karani.toml for one run.

.env.example is the template. Nothing in it is required to run a pass — with no variables set at all, karani uses an in-memory store and a local model.

Credentials

Variable	Needed for
`DATABASE_URL`	Keeping state between runs. Any Postgres DSN.
`OPENROUTER_API_KEY`	`provider = "openrouter"`
`OPENAI_API_KEY`	`provider = "openai"` against the API itself
`ANTHROPIC_API_KEY`	`provider = "anthropic"`
`SLACK_BOT_TOKEN`	Pushing to Slack (`xoxb-...`)
`SLACK_APP_TOKEN`	The Socket Mode listener (`xapp-...`)
`SLACK_CHANNEL`	Which conversation to push to
`NOTION_TOKEN`	The Notion mirror
`GMAIL_CLIENT_ID`	The watcher’s OAuth client
`GMAIL_CLIENT_SECRET`	The same
`GMAIL_REFRESH_TOKEN`	Reading the mailbox, read-only
`MINIO_ACCESS_KEY`, `MINIO_SECRET_KEY`	Object storage for the documents
`MEM0_PG_URL`	The vector store, separate from `DATABASE_URL` on purpose

A key for a provider you do not use is not needed, and its absence is never an error: with no key for the configured provider, karani falls back to the local model and says so once (ADR 0019).

Overrides

These shadow a karani.toml value for one run. Use them for experiments; put the answer in the file.

Model routing

Variable	Default	Notes
`QUAL_PROVIDER`	from `[llm.qualify]`	Any registered provider name
`QUAL_MODEL`	from `[llm.qualify]`	Provider-specific slug
`QUAL_REASONING_EFFORT`	`high`	`low`, `medium`, `high`
`QUAL_MAX_TOKENS`	`8000`	Completion cap, including reasoning tokens
`QUAL_TIMEOUT_SECONDS`	`180`	Thinking models are slow
`LOCAL_LLM_MODEL`	from `[llm.*]`	For `provider = "local"`
`RESUME_PATH`	`data/resume.md`	Which resume to judge against

Filtering

Variable	Default	Notes
`MIN_COMP_USD`	`160000`	Hard-fail comp floor where disclosed
`TARGET_COMP_USD`	`220000`	Score bonus threshold
`STALE_JOB_DAYS`	`10`	Days unseen before a posting is closed

HTTP

Variable	Default	Notes
`HTTP_TIMEOUT`	`30`	Per request, seconds
`HTTP_CONCURRENCY`	`6`	Global cap
`HTTP_PER_HOST_CONCURRENCY`	`3`	Per host
`USER_AGENT`	`karani/...`	Sent on every request

Switches

These turn a behaviour off, and each one exists because the behaviour costs something — money, network, or both.

Variable	Effect
`KARANI_AUTO_PREP=off`	Stop a `screen`/`interview` status change from building a prep pack. Two model calls, unattended.
`KARANI_ATS_QUESTIONS=off`	Stop the console reading real application forms. It then uses the standard field set and touches no network.
`KARANI_INTEL=off`	Stop company-dossier probes.
`KARANI_MEMORY`	`off`, `basic` (default) or `mem0`.
`KARANI_PASS_DEADLINE_SECONDS`	The hard ceiling on one hunt pass. Default 3000, clamped below the hourly tick (ADR 0025).

The test suite sets KARANI_AUTO_PREP and KARANI_INTEL off for the same reason it strips every credential: a deterministic test must not spend money or reach the network.

Ollama

Not karani’s variables, but karani’s behaviour depends on them, and getting them wrong is the single most expensive mistake available:

Variable	Set it to	Why
`OLLAMA_CONTEXT_LENGTH`	`16384`	The 4096 default silently truncates a qualification prompt
`OLLAMA_NUM_PARALLEL`	`1`	More than one slot splits the context window, quietly undoing the above

Infrastructure has the full story, including why launchctl setenv does not survive a reboot and what karani installs to work around it.

Resolution order

Loosest first:

Schema defaults
karani.toml
An environment variable
A command-line flag


karani config check   # what it resolved to, per task