Providers, Proxy and AG-UI Reference

Options for model accounts, GitHub Copilot sign-in, the local LLM proxy and the AG-UI server. For the explanations see GitHub Copilot Auth, Local LLM Proxy and AG-UI for Web Apps; for the step-by-step see the guide Use Pando from your editor and other apps.

Provider accounts

Declare only the providers you use. A section left in place with an empty key makes Pando offer an account that cannot work.

[Providers.anthropic]
APIKey   = ''
BaseURL  = ''
Disabled = false

[Providers.openai]
APIKey   = ''
Disabled = false

[Providers.copilot]
Disabled = false

[Providers.ollama]
BaseURL  = 'http://localhost:11434'
Disabled = false

[Providers.llama-cpp]
BaseURL  = 'http://localhost:8080'
Disabled = false

Models are written <provider>.<model>, for example copilot.gpt-4.1.

Prices and context sizes that a provider does not report are completed from models.dev, downloaded once and kept for 24 hours in ~/.pando_modelsdev.json. To never contact it:

[ModelsDev]
Enabled = false

GitHub Copilot

pando auth copilot login                                   # opens the browser
pando auth copilot login --enterprise-url https://github.mycompany.com
pando auth copilot login --no-browser                      # print the link and code
pando auth copilot status
pando auth copilot logout

Nothing else to configure: the sign-in is kept in your Pando profile.

Copilot plan	Models you get
Free	A limited set
Pro	GPT-4, GPT-4o
Pro+	A wider set
Business / Enterprise	Your organisation’s models, including the ones it added with its own keys (BYOK)

Organisation models are named with the organisation in front:

pando --model copilot.gpt-4.1 -p "Explain this code"
pando --model 'copilot.myorg/OpenRouter/some-model' -p "Explain this code"

Local LLM proxy

pando llm-proxy                       # localhost, port 11434
pando llm-proxy --port 8080
pando llm-proxy --host 0.0.0.0
pando llm-proxy --api-key mysecretkey   # ask clients for this key
pando llm-proxy --debug

What the other tool asks for	What to give it
API base URL	`http://localhost:11434/v1` (or the address the proxy prints when it starts)
API key	Any text, unless you started the proxy with `--api-key`

The proxy speaks the OpenAI format, and Anthropic’s /v1/messages too.

AG-UI server

pando agui-serve --port 8090 --allow-origin http://localhost:3000   # on its own (recommended)
pando serve --agui-port 8090                                        # next to the Web UI
pando agui-serve --print-token                                      # show the stored token again

Address for your page: http://localhost:8090/api/v1/agui/coder, with the header Authorization: Bearer <token>.

Flag	What it does
`--agent`	Agent to offer (repeat for several)
`--allow-origin`	Web address allowed to connect (repeat for several)
`--persona`	Persona added to every conversation
`--cwd`	Project folder to work in
`--host`, `--port`	Where to listen (`localhost`, `8090`)
`--token`, `--token-file`	Choose the token yourself. Also read from the `PANDO_AGUI_TOKEN` environment variable
`--no-token`	No token at all
`--auto-approve`	Approve the agent’s actions without asking the page
`--no-tls`, `--tls-cert`, `--tls-key`	Plain HTTP, or your own certificate

When you give no token, Pando creates one, prints it once and stores it in ~/.config/pando/agui-token for the next starts.

[AGUI]
Enabled        = true
Port           = 8090
Host           = 'localhost'
Agents         = ['coder']
AllowedOrigins = ['http://localhost:3000']   # empty means no browser may connect
RequireToken   = true
FrontendTools  = true
HumanInTheLoop = true

[AGUI.Profiles.reviewer]
Base    = 'coder'
Model   = 'anthropic.claude-sonnet-4'
Persona = 'code-reviewer'

Client library (TypeScript SDK):

import { PandoAguiClient } from '@pando-ai/sdk/agui';

const client = new PandoAguiClient({ baseUrl: 'http://localhost:8090', token });
for await (const event of client.run({ prompt: 'Summarise the repo' })) {
  if (event.type === 'TEXT_MESSAGE_CONTENT') process.stdout.write(event.delta);
}

A complete Next.js example with chat, a status board, a page tool and in-page approvals is in the Pando repository under examples/copilotkit.

Several Pando windows at once (IPC)

Nothing to configure: it sets itself up when more than one Pando is running. Internal defaults: heartbeat every 5 s, a window is considered gone after 15 s, a deeper check every 60 s. The list of running windows is kept in /tmp/pando-instances/.

To let windows hand work to each other:

[Mesnada.Delegation]
AllowExternalWarmTargets = true   # the one that asks
AcceptDelegations        = true   # the one that accepts

Message between windows	What it does
`state.sync`	Full copy of the current state
`session.list`	List sessions
`session.activate`	Open a session
`message.send`	Send a message
`session.interrupt`	Stop the model’s answer in progress
`instance.ping`	Check a window is alive and what it can do
`delegation.run`	Run a handed-over task
`delegation.cancel`	Cancel it
`delegation.status`	Ask how it is going
`db.compact`	Tidy the database (always done by the main window)

What the other windows hear live: sessions created, opened and deleted, new messages, the model’s answer as it is written, and tools starting and finishing.