Troubleshooting S.E.N.T.R.I

Connecting a model

Claude Code or Codex does not connect

S.E.N.T.R.I drives the CLI in the background and uses its sign-in, so the CLI has to be installed and signed in on this machine first. Open a terminal, run claude for Claude Code or codex for Codex CLI, sign in once, then try again. For Claude Code, whatever the CLI is set up for — a Claude subscription, an API key, Amazon Bedrock or a LiteLLM gateway — is what the agent uses.

OpenRouter requests fail

Check the key in the S.E.N.T.R.I settings, then the budget meter in the chat header. Once the monthly cap is used up, requests fail until it resets or you raise the limit on openrouter.ai. If only one model fails, set reasoning to Off — not every model accepts every effort level — or pick another model from the Recommended list.

Ollama does not answer

Make sure Ollama is running and that the endpoint in the S.E.N.T.R.I settings matches it, by default http://localhost:11434. The model you pick must already be pulled (ollama pull followed by its name). A large model on a small machine can take a long time to answer; try a smaller one.

While it works

It stops and waits

A skill marked approval-required pauses the run until you press Accept or Decline on the card in the chat, or answer it on WhatsApp. Scroll back to find the card. The agent cannot press it for you, not even with Computer Use.

It is going the wrong way

Press Esc to interrupt a running prompt, then say what you want instead. For a clean start, press New chat or type /newchat.

A skill is missing, or it says it cannot do something

Only enabled skills are offered. Switch the skill on in Settings and it appears in the Available Skills panel beside the chat. GitHub, Cloudflare, Google Search Console and Stripe also need their account connected in Settings first.

Every Coder call fails

Coder only works inside the working folder, and until one is picked every call fails. Set it with the folder icon in the chat header; each agent in the Agent Workspace has its own. Paths outside the folder are refused on purpose. Coder also refuses to edit a file it has not read in the conversation. The agent normally reads it first, so if you see that, ask it to read the file and try again.

It will not touch a server

The agent only sees the hosts in the current scope. Change the scope in the chat header or with /scope. Working on remote hosts over WinRM or SSH needs the Commander plan; on the S.E.N.T.R.I plan the agent works on this machine only.

Nothing shows on screen when it uses the browser

Browser Use runs headless on purpose. Ask S.E.N.T.R.I to show the Playwright dashboard to watch or take over, or to open a visible browser window. It cannot open file:// addresses, so to look at a local page or image, have it serve the folder over http://127.0.0.1 first.

The mouse pointer moves by itself

That is Computer Use. A real click borrows the pointer for a moment and puts it back, and a hover leaves it where it landed so a tooltip can be read. Keep your hands off the mouse until the step finishes.

Scheduled prompts

A scheduled prompt did not run

Check the state of the row in Schedules. Expired means a one-time schedule missed its moment by more than five minutes. Disabled means it is not registered: deleting its Windows task yourself disables the schedule instead of recreating it. A row saved before you picked a model asks to be saved again. Scheduled prompts also run one at a time, so one that comes due while another is working waits its turn.

It stopped before finishing

The one limit on an unattended run is the max runtime under Settings → S.E.N.T.R.I → Schedules: 60 minutes by default, and 0 for as long as it takes. It is checked between turns, so the step in flight always completes. Raise it for long jobs, and read the transcript in Logs\Sentri to see how far it got.

It did not use a skill it needed

Nobody is there to approve anything at 04:30, so the Allowed skills picker on the row is the approval: a skill that is not ticked is not offered to the run. Edit the row and tick it. An empty picker offers every enabled skill instead. See Schedules.

Chat integrations

WhatsApp does not answer

Open Settings, WhatsApp and check the status. If it is not connected, scan the QR code again. Write in the chat with your own number, which is the conversation S.E.N.T.R.I answers, and look for the reply in the main chat as well as on your phone.

Slack or Discord ignores a message

S.E.N.T.R.I answers direct messages and @sentri mentions, so a channel message without the mention is not for it. It also has to be a member of the channel. The answer comes from the agent bound to that channel in the Agent Workspace, not from the main chat.

Voice

The microphone picks nothing up

Push to talk is Shift+Space since v5.2.3. Check that Windows lets desktop apps use the microphone, under Settings → Privacy & security → Microphone. If dictation in your language comes out wrong, switch the speech-to-text model: Nemotron 3.5 ASR covers 40 languages.

There is no spoken reply

Only spoken prompts are answered out loud. Check that the speaker button in the prompt line is not muted. The voice comes from Grok Voice TTS through OpenRouter, so it needs a working OpenRouter key.

Memory and history

It forgot something it knew

Memory is kept per work folder. If the chat is on a different working folder, it sees a different memory: switch back, or tell it again. If it remembers something wrong, say so and it rewrites that memory. See Memory.

The history view is empty or slow to open

An empty view after a search and a freeze on first launch were fixed in v5.2.3 and v5.2.4. Whatever the view shows, the transcripts are always on disk under Logs\Sentri. See Chat History.

Fixed in recent versions

If you see one of these, updating ServerEngine is the fix.

Before you contact support

• Note the version you are on, and whether the problem is in the main chat, an agent in the Agent Workspace or a scheduled prompt.
• Open the Logs tab and press Captured Errors for the last error messages.
• Keep the transcript: Logs\Sentri\Chat-<USERNAME>-yyyy-MM-dd.txt for a chat, or the run file for a scheduled prompt. Transcripts are masked by default, but read one before you send it anywhere.