Using Evener with Ollama

Ollama lets you run open-weight LLMs locally. Evener's ollama provider speaks the openai-chat protocol (the same Chat Completions wire format every OpenAI-compatible vendor uses) with auth = "optional-bearer" and a host-normalizing rule, host_rule = "ollama-host", that turns OLLAMA_HOST/OLLAMA_BASE_URL into a base URL. Run the server on the Evener execution host or an endpoint reachable from it. A browser running on your laptop does not make localhost refer to that laptop from a remote session.

Quick start

1. Install Ollama — https://ollama.com/download

2. Start the daemon (the macOS/Windows app does this automatically; on Linux run ollama serve)

3. Pull a model that supports tool calling — Evener relies on native function-calling, so the model has to support it:

   ollama pull llama3.1:8b

4. Check the local endpoint, from the host that will run Evener:

   curl -fsS http://localhost:11434/v1/models

   Confirm the pulled model ID appears. This checks listing, not tool-calling quality. Configure the real context window as described below before a long task.

5. Run Evener from a project directory containing a README:

   evener --sandbox read-only --model ollama/llama3.1:8b "summarize the README"

   No API key, no env vars, nothing else needed if Ollama is at the default http://localhost:11434. Set OLLAMA_HOST or OLLAMA_BASE_URL to point at a non-default endpoint (see below).

   ollama has no curated default_model. Use the full ollama/<model> reference even if it is your only connection. To make a local model the default deliberately, merge this complete document into providers.toml, substituting a model you have pulled:

   default = "ollama"
   
   [providers.ollama]
   default_model = "llama3.1:8b"

   Setting only default = "ollama" does not invent a default model. Other credentialed instances with defaults can win automatic ranking when no explicit default is set.

How it works

protocol = "openai-chat" — the same Chat Completions protocol other implicit providers use. auth = "optional-bearer" sends a bearer token when a key resolves, and nothing otherwise. host_rule = "ollama-host" is the one normalizer described below. Models, tool calls, multimodal images (for models that support them), and /v1/models listing all flow through that protocol's normal code path.

Environment variables

Variable	Purpose	Default
OLLAMA_BASE_URL	Full URL including /v1	http://localhost:11434/v1
OLLAMA_HOST	Ollama's canonical env var (host, host:port, or full URL); used when OLLAMA_BASE_URL is unset	unset
OLLAMA_API_KEY	API key for authenticated proxies or Ollama Cloud	unset

Resolution order for the base URL:

1. If OLLAMA_BASE_URL is set, it wins (used as-is, trailing slash stripped).
2. Otherwise, if OLLAMA_HOST is set, it's normalized:
   • bare host (ollama.local) → http://ollama.local:11434/v1
   • the bare host ollama.com (Ollama Cloud) is the one exception to that rule → https://ollama.com:443/v1 (https, port 443, not the usual 11434)
   • host:port (192.168.1.5:11434) → http://192.168.1.5:11434/v1
   • localhost → http://localhost:11434/v1
   • a bare IPv6 literal (::1) → http://[::1]:11434/v1
   • full URL (https://ollama.example.com) → https://ollama.example.com/v1
   • a URL whose path already ends in /v1 (e.g. https://proxy/ollama/v1) is preserved verbatim
3. Otherwise the default http://localhost:11434/v1 is used.

Examples

Local default

OLLAMA_HOST=localhost evener --model ollama/llama3.1:8b "what does main.go do?"

Remote Ollama on your LAN

This sends prompts and tool results over the network. Use a trusted private network or an authenticated encrypted proxy; do not expose an unauthenticated Ollama listener publicly.

OLLAMA_HOST=192.168.1.5:11434 \
  evener --sandbox workspace-write --model ollama/llama3.1:8b "fix the failing test"

Ollama behind an authenticated proxy

OLLAMA_BASE_URL=https://ollama.example.com/v1 \
OLLAMA_API_KEY="$MY_PROXY_TOKEN" \
  evener --model ollama/llama3.1:70b "task"

One-shot calls with llmcall

llmcall is a separate single-call executable, not an evener subcommand. See its build and installation note.

llmcall --provider ollama --model llama3.1:8b "Write a haiku about goroutines."

Listing locally available models

curl -s http://localhost:11434/v1/models | jq -r '.data[].id'

Choosing a model

Evener drives every step through native tool calling. Models without function-calling support will fail to make progress.

Check the installed model's tool-calling support in Ollama's model metadata. Model tags, templates, quantization, and hardware affect results; a successful /v1/models listing does not prove that a model can complete a coding task. Start with a read-only task, inspect its tool calls, and verify its answer against the files.

Context length

When a live model's listing supplies no context length, Evener cannot infer its server-configured window. Live-only models on ollama and the protocol-only pseudo-providers use a provider-level default of context_window = 131072 unless live metadata or a user setting supplies a window. Evener does not use a bundled Ollama per-model table or strip tags to find a context-window default.

Pin the actual server-configured window on a model row in providers.toml. This complete example declares 8192 tokens; change it to match your server. It changes Evener's budget, not Ollama's num_ctx:

[providers.ollama.models."llama3.1*"]
context_window = 8192

If the server's window is smaller than Evener's budget, compaction can start too late to prevent a context-limit error. Check the configured window before a long task. For changes from older bundled defaults, see provider migration.

Troubleshooting

connection refused — the Ollama daemon isn't running, or it's bound to a different host/port. Check with curl http://localhost:11434/api/tags.

model 'X' not found — pull it first: ollama pull X.

Bare text instead of tool calls — the model isn't honoring tool schemas. Check its template and tool support, then try a model suited to tool use. Repeating the task without changing the model setup may repeat the failure.

Truncated long conversations — the model's num_ctx is smaller than the window Evener is budgeting against. Pin the real window with [providers.ollama.models."<glob>"] context_window = ...; see Context length above.

Slow responses — local inference is bound by your hardware. Use a quantized model (:q4_K_M etc.), a smaller model, or run Ollama on a machine with a GPU and point Evener at it via OLLAMA_HOST.

See also

• llm-providers.md — the registry: layers, instances, and how ollama fits the implicit-provider table.
• llm-provider-config-and-launch.md — credentials, OAuth, and how config reaches spawned sessions.
• Environment variables — the OLLAMA_HOST/OLLAMA_BASE_URL/OLLAMA_API_KEY env vars alongside every other provider's.