Run Obsidian Copilot on any catalog model.

Quick answer: Add Custom Model, four fields.

Install the Copilot plugin from Obsidian's community plugins if you have not, then open Copilot's settings and go to the Model tab. Click Add Custom Model and fill the form: the exact model id as the model name, provider set to 3rd party (openai format), Base URL https://api.apisrouter.com/v1, and your APIsRouter key. Use the Verify button before saving: it sends a live test request, so a bad id or key fails here instead of mid-conversation. Once verified, the model appears in Copilot's model dropdown, and switching between a Claude id and a DeepSeek id is a dropdown choice, not a reconfiguration.

Model name:  claude-sonnet-4-6
Provider:    3rd party (openai format)
Base URL:    https://api.apisrouter.com/v1
API Key:     sk-YOUR-APISROUTER-KEY

→ click Verify, then save

How Copilot uses the model.

Copilot for Obsidian (logancyang/obsidian-copilot, roughly 7.4K GitHub stars) is the most-installed AI plugin for the Obsidian note-taking app: chat with your notes in a sidebar, ask questions across the whole vault, and run commands over selected text. The plugin speaks the OpenAI chat-completions format to whatever Base URL each model entry declares, which is why any OpenAI-compatible endpoint slots in as a provider. The traffic shape is distinctive: vault Q&A stuffs retrieved note chunks into context, so requests are input-heavy, and a long research session over a large vault sends the same kind of context repeatedly. That makes the model choice a real economic decision: a volume-tier id for everyday note lookups and a flagship for hard synthesis is the pattern that fits, and both live in the same dropdown once configured. One mechanical caveat from the plugin's own guidance: Obsidian is an Electron app, so some providers trip CORS restrictions. Copilot ships a CORS workaround toggle on custom models; if a verified model errors when chatting, enable that toggle and check the developer console before assuming the endpoint is down.

Choosing models for a knowledge vault.

Vault Q&A resends retrieved context on every question, so input tokens dominate the bill. The per-model usage view in the gateway console shows exactly what a week of your vault workflow costs on each id, which makes the flagship-versus-volume decision empirical rather than aesthetic.

• claude-haiku-4-5-20251001 for everyday note Q&A: fast, low-rate, and strong enough for retrieval-grounded answers.
• claude-sonnet-4-6 for synthesis across many notes, writing assistance, and anything where answer quality is the point.
• gpt-5.4-mini as the GPT-family volume pick if your prompts and templates grew up on OpenAI behavior.
• gemini-3.5-flash when sessions ingest very long notes or transcripts at a mid-tier rate.
• deepseek-v4-flash as the floor-rate option for bulk operations like summarizing or tagging across a folder.

Pay-as-you-go · transparent per-model pricing

Selected models are priced below official list prices. Exact input, output, cache, and per-request prices are shown for each model.

Failure modes specific to Obsidian Copilot.

The plugin is updated frequently and its settings layout has moved before, so if a field name here does not match your version exactly, the shape stays the same: model id, openai-format provider, Base URL, key, Verify. The gateway half is provable with one curl command whenever you are unsure which side is failing.

• Verify fails immediately: the model name does not match a catalog id, or the key is wrong. Curl the /v1/models listing with the same key and copy the id exactly.
• Verified but chat errors: usually CORS inside Electron. Enable the CORS workaround toggle on the custom model entry and re-test; the developer console (Ctrl+Shift+I) shows the blocked request if that is the cause.
• Answers ignore your notes: that is retrieval configuration, not the model endpoint. Check Copilot's vault QA settings and index state before blaming the id.
• Embedding features failing while chat works: embedding models are configured separately in Copilot and are a different API surface. Verify any embedding id against the catalog before enabling vault-wide semantic search through the gateway.
• Base URL with a trailing /chat/completions: the plugin appends route paths itself. Give it the base ending in /v1 and nothing more.

Who routes Obsidian Copilot through a gateway.

• Note-takers who want Claude quality on vault questions without an Anthropic billing account, on a prepaid balance with a free start.
• Researchers running long Q&A sessions over large vaults, where input-heavy traffic makes per-token rates matter.
• Users switching between a volume id for lookups and a flagship for synthesis, both entries in the same model dropdown.
• People already routing editors and terminals through a gateway who want their notes app on the same key and usage log.
• Privacy-conscious users consolidating AI traffic through one endpoint they can audit per request in a console.

Verify the endpoint, then verify in the plugin.

Do the two-command check first: list the models with your key, then send one completion with the exact id you plan to add. When both pass, the plugin's Verify button should pass too; if it does not, the gap is in the form fields or CORS, which the developer console will show. Once chatting works, the APIsRouter console records each request with model, token counts, and spend. A vault session is many requests carrying note context, so the console, filtered to the key you use for Obsidian, is the honest record of what your second brain costs to consult.

curl -s https://api.apisrouter.com/v1/models \
  -H "Authorization: Bearer $APISROUTER_API_KEY" | head -50

curl -s https://api.apisrouter.com/v1/chat/completions \
  -H "Authorization: Bearer $APISROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"claude-haiku-4-5-20251001",
       "messages":[{"role":"user","content":"ping"}]}'

FAQ