By default, Shelley on exe.dev VMs runs on exe.dev's managed LLM gateway. These steps switch the Anthropic provider to your own API key ("BYOK"). After the switch, you confirm it from inside a VM and check that it applies to every VM you own.
How it works: the key is stored in exe.dev's LLM integration, not on your VMs. VMs reach models at
https://llm.int.exe.xyz, and exe.dev adds your key at the network edge. Code on a VM, including an agent, can use the key but can never read it.
References: Bring Your Own Key · LLM Integration · Attaching Integrations
Pick one of these options.
- Start a new conversation.
- Open the model drop-down and choose Add / Remove Models...
- Choose Anthropic and paste your key.
- Open https://exe.dev/integrations and click LLM.
- Edit the
llmintegration and set Anthropic to API Key. - Paste the key and, if you like, click Test.
- Make sure it's attached to
auto:all, which means every VM.
Use - so the key is read from stdin. That keeps it out of shell history and the process list:
printf '%s' "$ANTHROPIC_API_KEY" | \
ssh exe.dev integrations edit llm --anthropic=byok --anthropic-key=-Editing the existing llm integration keeps its attachments. Out of the box, that's auto:all.
If the key is in a .env file, pipe it in without ever displaying it:
grep -E '^ANTHROPIC_API_KEY=' .env | cut -d= -f2- | tr -d '\n' | \
ssh exe.dev integrations edit llm --anthropic=byok --anthropic-key=-To switch back to exe.dev's managed credentials:
ssh exe.dev integrations edit llm --anthropic=gatewayRun these on a VM.
Provider mode: Anthropic should report byok.
curl -s https://llm.int.exe.xyz/models.json | jq '.providers | map_values(.mode)'Expected output (other providers will still say managed unless you changed them too):
{
"anthropic": "byok",
"deepgram": "managed",
"fireworks": "managed",
"openai": "managed",
"xai": "managed"
}End-to-end call: this proves the key actually works. Model IDs use dashes and a provider prefix, for example anthropic/claude-haiku-5-5. List the available IDs with jq -r '.models[].id' on models.json.
curl -s https://llm.int.exe.xyz/v1/messages \
-H 'content-type: application/json' \
-H 'anthropic-version: 2023-06-01' \
-d '{"model":"anthropic/claude-haiku-5-5","max_tokens":5,"messages":[{"role":"user","content":"hi"}]}' \
| jq '{model, type, text: .content[0].text}'An HTTP 200 with "type": "message" means your key is being used. A 401 or 403 from upstream usually means the key is invalid or has no credit.
Do this from your local machine. A VM can't see account-wide integration settings.
# 1. The integration: look for anthropic=byok and an auto:all attachment.
# Stored secrets appear as ***.
ssh exe.dev integrations list --json
# 2. Your VMs
ssh exe.dev ls
# 3. Spot-check any VM (should print "byok")
ssh <vm>.exe.xyz "curl -s https://llm.int.exe.xyz/models.json | jq -r .providers.anthropic.mode"Optional loop over several VMs:
for vm in vm-one vm-two vm-three; do
printf '%-20s ' "$vm"
ssh "$vm.exe.xyz" "curl -s https://llm.int.exe.xyz/models.json | jq -r .providers.anthropic.mode"
doneThe setting covers every VM only if:
- The
llmintegration is attached toauto:all, not to onetag:orvm:. To fix it:ssh exe.dev integrations attach llm auto:all
- No other LLM integration attached to a VM changes what that VM sees. Look in
integrations listfor extra LLM-type entries attached to a tag or to individual VMs.
| Symptom | Likely cause / fix |
|---|---|
ssh exe.dev from a VM prints "Please complete registration" |
That VM isn't logged in to the exe.dev CLI. Run account-level commands from your local machine. |
curl https://reflection.int.exe.xyz/ says "integration not found or not attached" |
The default reflection integration was removed. Reinstall it with ssh exe.dev int add reflection --default. |
Model "..." is not in this exe.dev integration's model list |
Wrong model ID format. Use the ID from models.json, for example anthropic/claude-haiku-5-5. |
/v1/models lists no Claude models |
That endpoint is the OpenAI-compatible list. Use /models.json to see all providers. |
Anthropic shows managed on one VM |
llm isn't attached to that VM (see step 3), or a different LLM integration covers it. |
- Never paste the key into a command argument. Always pipe it in and use
--anthropic-key=-. integrations list --jsonmasks secrets as***. If the key is ever exposed, revoke it in the Anthropic Console and run step 1 again with a new key.