Models and tools¶
The composer's two buttons configure the agent for the current thread. This page is the full reference. The same controls appear in every chat, in the Agent column or a thread tab.
The model button¶
The model button always shows the active model. Tap it for:
- Favorites — models you starred in the full picker, one tap each.
- Other models — the full picker across every configured provider. Search by name, browse by vendor, and star models to build your favorites list.
- Thinking — reasoning effort, shown only when the selected model supports it. The levels offered come from the model itself.
- Max reply tokens — a cap on each reply's length. With PlebChat credits it also caps how much one agent step can hold from your balance, so a small cap lets a small balance run long agent loops.
Picking a model also makes it the default for your next new chat. Thinking and the reply cap belong to the thread you set them in.
The + button¶
Everything besides the model. Icons on the + button show what is active.
Tools¶
Each tool switches on per thread. A tool's own settings sit under it in the list.
| Tool | What it does | Needs |
|---|---|---|
| Web | Searches the web and reads pages; records sources for citation | An active web search provider: PayGate credits, or a Firecrawl, Tavily, or Exa key (Settings → Custom Providers) |
| Nostr | Full-text search over Nostr notes and long-form articles | Nothing |
| Python | Runs Python in a sandboxed in-browser interpreter, with numpy, pandas, and matplotlib. The chat's files are its working directory: it reads what Fetch saved and writes results and charts back as chat files. Stop a long run from its card | Nothing |
| Files | Reads, writes, and edits files attached to the thread | Nothing |
| Fetch | Reads a URL as raw text or data (JSON, CSV) and saves it to the chat's files | Nothing for sites that allow browser reads; sites that block them fall back to a paid relay fetch (1 sat) when you hold PlebChat credits |
The web tool's Search mode chooses how hard it searches: Balanced reads the best few pages; Deep Research decomposes the question, searches from several angles, and cross-checks. A Research agent mode is planned and not yet available.
Your tool picks also become the default for new chats.
Permissions¶
Costly or destructive calls — web searches and page fetches, file writes and edits — are Allowed by default. Set a tool to Ask each time to pause each such call on an approval card. This choice is per tool and lasts only for the current browser session; a reload returns to Allowed.
Attachments¶
Attach files or photos opens the file picker (Chat only). Photos require a model that declares image support; the attach row explains when the current model does not.
Personality¶
A personality prompt shapes who the agent is. Pick one per thread:
- PlebChat default — the stock persona.
- A prompt from your library. Deep Research, Analyst, and Socratic Tutor ship built in; write your own under Settings → Personality & Skills.
- Write your own — type a prompt for this thread only.
By default the personality is added after the stock PlebChat opener. Untick Include the PlebChat preamble to let it stand alone. Tool instructions are always added for whatever tools are on.
Skills¶
A skill is a written working method the agent can load on demand. Its name and description are shown to the agent; the full instructions load only when needed. Toggle your own skills per thread; some skills come bundled with a tool (Python and Files each carry one) and switch on with it.
Write skills under Settings → Personality & Skills.
What carries over¶
| Setting | Scope |
|---|---|
| Model | The thread; also seeds new chats (last used) |
| Thinking level, max reply tokens | The thread |
| Tools, search mode | The thread; also seed new chats |
| Personality, skills | The thread only |
| Ask each time | This browser session only |
Personality prompts and skills you create sync encrypted to your relays. Thread settings travel with the thread's own sync policy.