Ollama Web Search

Ollama web search, called securely

When to use

• User wants live web results (not just a plain web_fetch of one known URL) and mentions Ollama, or already has an Ollama account/CLI installed.
• Any time a task needs a secret (API key, token) passed into a shell command but must not appear in the visible conversation or in any file I create.

The two Ollama auth mechanisms — don't confuse them

• ollama signin / the id_ed25519 keypair in ~/.ollama/ — SSH-style identity used for the model registry (push/pull/publish). Being "signed in" this way grants nothing for the REST API below.
• Web Search / Web Fetch REST API (https://ollama.com/api/web_search, .../api/web_fetch) — needs a separate bearer token minted at https://ollama.com/settings/keys, sent as Authorization: Bearer <key>. ~/.ollama/config.json does not store this; there is no local daemon proxy for it either (127.0.0.1:11434/api/web_search 404s — the local Ollama server only serves inference endpoints).

Secure key handling protocol

Never ask the user to paste a secret into the chat, and never write a command whose literal text contains the key. pi's bash tool logs record command text, so the key must never appear in the command string itself.

1. Ask the user to create the key file in their own terminal, outside the conversation:

   printf '%s' 'THE_KEY' > ~/.ollama_web_search_key
   chmod 600 ~/.ollama_web_search_key
   

2. Once they confirm it exists, verify presence/permissions only — never cat the file to check it:

   test -f ~/.ollama_web_search_key && stat -f "%A %N" ~/.ollama_web_search_key
   

3. Use command substitution inline so the key only lives transiently in the child process's environment, never in the command text or any tool-call log:

   curl -s https://ollama.com/api/web_search \
     -H "Authorization: Bearer $(cat ~/.ollama_web_search_key)" \
     -H "Content-Type: application/json" \
     -d '{"query":"...", "max_results": 8}'
   

   • No -v/verbose curl flags (they can echo request headers back).
   • Only the JSON response body — public search results — re-enters context; the key itself never does.
4. If the response is large, the bash tool may save full output to a temp file — parse it with python3 -c "import json..." or jq, and prefer jq/python3 filters over dumping raw output. The temp file holds public search results, not the secret, but there's no reason to leave litter if you can clean it up.
5. Leave the key file for reuse, or let the user rm it — that's their call, not something to do unprompted.

API shape reference

• POST /api/web_search — body {"query": str, "max_results"?: int (≤10)} → {"results": [{"title", "url", "content"}]}.
• POST /api/web_fetch — body {"url": str} → {"title", "content", "links"}.