Environment variables

elelem has no config file. Settings that are not command-line options are environment variables, and most belong to a plugin. This is every variable read by core and the plugin gems, found by searching each repository for ENV and getenv.

“Required” means the provider stops with an error when the variable is missing. “Build” means the variable is read when the gem’s native extension compiles, during gem install, not at run time.

Variable	Read by	Default	What it does
`ELELEM_LOG_LEVEL`	`elelem`, `elelem-llama`	`warn`	Log level for messages on stderr: `debug`, `info`, `warn`, `error`, or `fatal`. `elelem-llama` applies it to llama.cpp’s own log too.
`ELELEM_CMD_TIMEOUT`	`elelem-builtins`	none	Seconds before an `execute` command is stopped with TERM, then KILL half a second later. The result notes the timeout and reports exit status 124. Unset means no limit.
`XDG_STATE_HOME`	`elelem-builtins`	`~/.local/state`	Session transcripts are written to `elelem/sessions/` under this directory.
`ANTHROPIC_API_KEY`	`elelem-anthropic`	required for `-p anthropic`	API key, sent as `x-api-key`.
`ANTHROPIC_MODEL`	`elelem-anthropic`	`claude-opus-4-5-20250514`	Model for `-p anthropic` when `--model` is not given.
`VERTEX_MODEL`	`elelem-anthropic`	`claude-opus-4-5@20251101`	Model for `-p vertex` when `--model` is not given.
`GOOGLE_CLOUD_PROJECT`	`elelem-anthropic`	required for `-p vertex`	Google Cloud project for Vertex AI. The access token comes from `gcloud auth application-default print-access-token`.
`GOOGLE_CLOUD_REGION`	`elelem-anthropic`	`us-east5`	Vertex AI region.
`OPENAI_API_KEY`	`elelem-openai`	required	API key, sent as a bearer token.
`OPENAI_MODEL`	`elelem-openai`	`gpt-4o`	Model when `--model` is not given.
`OLLAMA_HOST`	`elelem-ollama`	`localhost:11434`	Host and port of the Ollama server.
`OLLAMA_MODEL`	`elelem-ollama`	`gpt-oss:latest`	Model when `--model` is not given.
`OLLAMA_THINK`	`elelem-ollama`	none	Sent as the `think` field of each request.
`OLLAMA_KEEP_ALIVE`	`elelem-ollama`	none	Sent as the `keep_alive` field of each request.
`ELELEM_GGUF_MODEL`	`elelem-llama`, `elelem-ggml`	`~/.agents/models/default.gguf`, or the only `.gguf` file there	Path to the GGUF model. For `elelem-llama`, `--model` takes precedence.
`ELELEM_GGUF_N_CTX`	`elelem-llama`	`0`	Context size in tokens. `0` uses the model’s training context, capped at 65,536. If that does not fit in memory, it halves the size down to 4,096.
`ELELEM_GGUF_N_GPU_LAYERS`	`elelem-llama`	`999` or `0`	Layers to put on the GPU. Default is `999` (all) on a Vulkan, CUDA, or Apple Silicon Metal build when the model file is under 80% of free GPU memory, otherwise `0`.
`ELELEM_GGUF_MAX_TOKENS`	`elelem-llama`, `elelem-ggml`	`0` (llama), `256` (ggml)	Most tokens to generate per reply. For `elelem-llama`, `0` means the rest of the context.
`ELELEM_GGUF_TEMP`	`elelem-llama`	`0.7`	Sampling temperature. `0` or below means greedy decoding.
`ELELEM_GGUF_SEED`	`elelem-llama`	`-1`	Sampling seed.
`ELELEM_TOP_K`	`elelem-llama`	`40`	Top-k sampling. Ignored when decoding is greedy.
`ELELEM_TOP_P`	`elelem-llama`	`0.95`	Top-p sampling. Ignored when decoding is greedy.
`ELELEM_REPEAT_LAST_N`	`elelem-llama`	`0`	Tokens to look back for the repeat penalty. `0` turns the penalty off.
`ELELEM_REPEAT_PENALTY`	`elelem-llama`	`1.0`	Repeat penalty. `1.0` is no penalty.
`ELELEM_KV_CACHE_TYPE`	`elelem-llama`	`q8_0`	KV cache type: `f16`, `f32`, `q4_0`, or `q8_0`. Anything else means `q8_0`.
`ELELEM_LLAMA_BACKEND`	`elelem-llama` (build)	`auto`	GPU backend to compile: `cpu`, `vulkan`, `cuda`, `hip`, or `metal`. `auto` picks Metal on macOS, then CUDA if `nvcc` is found, Vulkan if `glslc` and the Vulkan loader are found, HIP if `hipcc` and a GPU target are found, else CPU.
`ELELEM_HIP_ARCH`	`elelem-llama` (build)	detected, else `gfx1100`	AMD GPU target for a HIP build.
`ELELEM_CANDLE_MODEL`	`elelem-candle`	required	A safetensors directory or a Hugging Face repo id.
`ELELEM_CANDLE_DEVICE`	`elelem-candle`	`cpu`	Device name passed to Candle.
`ELELEM_LLAMA2_CHECKPOINT`	`elelem-llama2`	required	Path to a llama2.c `.bin` checkpoint.
`ELELEM_LLAMA2_TOKENIZER`	`elelem-llama2`	the gem’s bundled `tokenizer.bin`	Path to the tokenizer.
`ELELEM_LLAMA2_STEPS`	`elelem-llama2`	`0`	Most tokens to generate. `0` means the model’s full sequence length.
`ELELEM_LLAMA2_TEMP`	`elelem-llama2`	`1.0`	Sampling temperature.
`ELELEM_LLAMA2_TOPP`	`elelem-llama2`	`0.9`	Top-p sampling.
`ELELEM_LLAMA2_SEED`	`elelem-llama2`	`-1`	Sampling seed.
`ELELEM_SD_IMAGE_MODEL`	`elelem-diffusion`	none	Stable Diffusion model for the `image` tool. The tool returns an error while it is unset.
`ELELEM_SD_VIDEO_MODEL`	`elelem-diffusion`	none	Diffusion model for the `video` tool. Required by that tool.
`ELELEM_SD_VIDEO_VAE`	`elelem-diffusion`	none	VAE for the `video` tool. Required by that tool.
`ELELEM_SD_VIDEO_T5XXL`	`elelem-diffusion`	none	T5-XXL text encoder for the `video` tool. Required by that tool.
`ELELEM_SD_ESRGAN_MODEL`	`elelem-diffusion`	none	ESRGAN model for the `upscale` tool. Required by that tool.
`ELELEM_PROVIDER`	`elelem-server`, `elelem-evals`	`stub` (server), `llama` (evals)	Provider when `-p` is not given. The `elelem` command does not read it.
Any `${NAME}` in `mcp.json`	`elelem-mcp`	required when referenced	Substituted into a server’s `env` and `headers`. A missing variable is an error for that server.
`CC`	`elelem-ggml`, `elelem-llama2` (build)	`cc`	C compiler.
`CXX`	`elelem-llama`, `elelem-llama2`, `elelem-diffusion` (build)	`c++`	C++ compiler.

The build scripts of elelem-llama and elelem-ratatui also search PATH for build tools: nvcc, glslc, hipcc, and vulkaninfo for llama, cargo for ratatui. Paths under ~ are resolved through HOME.

Setting them #

For one command:

ELELEM_LOG_LEVEL=debug elelem ask -p ollama "hello"

For every session, export them from your shell profile:

export ANTHROPIC_API_KEY=your-key
export ELELEM_CMD_TIMEOUT=120

--model on the command line wins over the provider’s model variable for anthropic, vertex, openai, ollama, and llama.