LLM Providers
Configure Ollama, Anthropic or any OpenAI-compatible host; nothing in the deterministic core needs one.
An LLM is optional and is used only by opt-in or read-side features: ekos ask, SQL semantic naming, document semantics, architecture role classification, [llm-description], [semantics] llm-definitions, ekos check explain, ekos migrate reconstruct, docs generate --prose, and in extensions binary reconstruction, Pentaho translation and lakehouse requirements. The compiler pipeline itself makes no model call unless one of these is turned on; on main, every run log ends with the count.
A
[llm]section in a project'sekos.tomlturns on SQL semantic naming atekos recover: one model call per schema file. Leave[llm]out of a project that should compile without one.
[llm]
provider = "ollama" # local, no key
model = "llama3:latest"
context-window = 8192 # sent as num_ctx
[llm]
provider = "anthropic"
model = "…"
api-key-env = "ANTHROPIC_API_KEY"
[llm]
provider = "openai" # any OpenAI-compatible Chat Completions host
base-url = "https://opencode.ai/zen/v1"
model = "deepseek-v4-flash"
api-key-env = "OPENCODE_API_KEY"
On main, provider = "zen" reaches OpenCode Zen with one key (OPENCODE_API_KEY by default),
choosing the wire format — Messages, Responses or Chat Completions — from the model.
| Key | Meaning |
|---|---|
provider |
ollama, anthropic, openai or zen |
model |
model name |
api-key-env |
name of the env var holding the key |
base-url |
openai (and zen) only; falls back to OPENAI_BASE_URL |
context-window |
ollama only; falls back to OLLAMA_NUM_CTX, then 8192 |
max-tokens |
ceiling for compiler-pass calls; leave unset to let each pass size its budget |
Routes per task
[llm]
provider = "ollama"
model = "llama3"
budget-usd = 1.0 # refuse a metered call that could take one command past $1
[llm.tasks.ask] # every field not given is inherited from [llm]
provider = "zen"
model = "claude-haiku-5-5"
fallback = ["<model>", …] lists models tried in order when a route's model answers 429 or 5xx;
like every field, a task inherits it from [llm] unless it sets its own.
Tasks: default, sql-naming, doc-semantics, description, semantics, session-capture,
ask, architecture, reconstruct, nl-sql, docs-prose, marketing, translate,
requirements, xref-triage. Other [llm] keys: prices (per million tokens, overriding the
built-in table), prompt-cache, reasoning-effort, allow-data-collection (off: models whose
provider may keep or train on input are refused).
ekos llm routes # each task's provider, models, wire format and price
ekos llm usage --since 2026-10-01 # calls that reached a model: tokens and cost by task and model
ekos llm probe # one tiny, uncached request per model — this is a model call
The read side: [ai]
[ai] tunes ekos ask: model, max-matches, neighborhood-depth, max-tokens, max-context-chars, source-text-top-k, system-prompt (classic path), reason-system-prompt. Reasoning models spend tokens on hidden reasoning; raise max-tokens (8192 is a good start) if answers truncate.
Cost control
Responses are cached in .ekos/llm-cache/. Steps that make many calls ([llm-description], [document-semantics]) are off by default, show an estimate and ask before spending, and accept --yes for automation. Commands that call a model on request — ekos check explain, ekos migrate reconstruct — refuse a model that is not a local Ollama one without --allow-metered. ekos doctor shows the provider, endpoint and whether the key variable is set; ekos llm usage shows what was actually spent.