Providers

A provider is what an agent thinks with: an official CLI on your computer, a model API, or the free Demo. The provider decides what an agent can do. A Claude Code agent can edit files in the project folder, while an API agent works with text and Agentbits’ own tools only. You choose a default provider and can give any agent a different one.

What each provider can do

	Claude Code CLI	Codex CLI	Anthropic API	OpenAI API	Demo
You need	Claude Code installed and signed in	Codex CLI installed and signed in	An Anthropic API key	An OpenAI API key	Nothing
Project folder	Required	Required	Optional	Optional	Optional
Read and edit files in the project folder	Yes	Yes, in its sandbox	No	No	Simulated
Run commands	Yes	Yes, in its sandbox	No	No	Simulated
Web access	Yes	Yes	No	No	No
Connections (SSH, databases)	Yes	Not yet	Yes	Yes	Yes
Attached files	Yes	In the prompt only	Yes	Yes	Yes
Images	Yes	Yes	Yes, plus scanned PDFs	Yes (PDFs not sent natively)	No
Skills and apps (MCP)	Yes	Yes	Yes	Yes	Yes
Token usage	Full, with cost in $	Total only	Full, no cost	Full, no cost	Estimated

Everything in the table still depends on the agent’s permissions. For example, a Claude Code agent edits files only with Write files on, runs commands only with Terminal on, and reaches the web only with Network on. Require approval works with every provider.

Claude Code CLI

Agentbits runs the official claude CLI in the background. File and terminal tools are turned on according to the permission matrix.

1. Install Claude Code on your computer.

2. Sign in once in a terminal. You can use Terminal → Shells inside Agentbits:

   claude auth login

3. In Agentbits, open Settings → Providers → Add provider and choose Claude Code CLI. The form shows ● installed if it found claude. Leave CLI path empty to find it on your PATH, or enter the full path.

4. Pick a Model and click Save and test. The test checks the CLI version and that you are signed in.

Notes:

• The project must have a folder. Agents start in it, and file edits outside it (and any extra allowed folders) are refused.
• Agentbits starts Claude Code with your project’s own .claude settings, hooks and .mcp.json ignored, so a repository can’t change how the agent runs.
• Besides the listed models you can use aliases such as sonnet or opus. An alias always uses the newest model of that family on your account.

Codex CLI

Agentbits runs OpenAI’s codex exec command in sandbox mode.

1. Install Codex CLI.

2. Sign in once in a terminal:

   codex login

3. In Settings → Providers → Add provider, choose Codex CLI, check that it is found, pick a model and click Save and test.

Notes:

• Codex always runs commands in its own sandbox. It writes only inside the project folder, and gets network access only when the agent has the Network permission.
• A Codex agent needs Read files. Without it, the agent doesn’t run.
• Codex agents can’t use connections yet. Attached files go into the prompt as far as they fit, and images are passed to the CLI. Extra allowed folders are not supported.
• Codex reports only a total token count, without cost or an input/output split.

Note: Some parts of the Codex integration (apps and images) have so far been tested only against a simulated CLI. If something doesn’t work, please report it.

Anthropic API and OpenAI API

These call the model directly with your API key: the Anthropic Messages API, or the OpenAI Chat Completions API.

1. Get an API key from Anthropic or OpenAI.
2. In Settings → Providers → Add provider, choose the provider, paste the API key, pick a Model and click Save and test.

The key is encrypted with your operating system’s keychain and stored locally. It is never sent back to the interface, to agents or to logs. To change a key, delete the provider and add it again.

API agents have no file system or terminal access. They can’t change files in the project folder or run commands. They work with text plus Agentbits’ own tools: the connections they are allowed to use, the files shared with them, saving new text files, skills and apps. For code changes, use a CLI provider.

When you enter a key, the model list also shows Other models on your account. OpenAI’s GPT-6 models are not listed, because they can’t use tools through Chat Completions. Use them through Codex CLI instead.

Demo (simulation)

Demo needs no key and calls no model. Agents behave as if they were really working: the lead splits your request into tasks, agents “use tools”, write reports and go to review. It never touches your files. When a task is about data or servers, Demo makes real, harmless calls through your connections, so you can see permissions, approvals and the audit log at work.

Use Demo to learn the app and try templates. It doesn’t do real work. If you have no providers set up at all, agents run in demo mode.

Mix providers and models

The default provider is used for every agent that doesn’t have its own. To change it, click Make default next to a provider in Settings → Providers.

To give one agent a different provider or model:

1. Open Organization → Agents and select the agent.
2. In the Model section, choose a Provider and a Model.

Provider’s model uses whatever model the provider is set to. If you switch the agent to a different kind of provider, its model is reset. Models are picked from a list. For a model that isn’t listed yet, choose Custom model name… and type its exact name.

For example, developers on Claude Code CLI can work in the project folder while an API agent writes up research or analyzes attached documents.

If sign-in expires

If a CLI session expires or a key stops working, Agentbits posts a message in the project’s #general, marks the agent with an error and pauses the whole team. Sign in again (claude auth login or codex login) or replace the key, test the provider in Settings → Providers, then click Resume in the top bar.

What is not supported

• Agentbits does not include, install or sign in to the CLIs. You install them and sign in yourself.
• You can’t sign in with a Claude account or subscription inside the app. Use an API key or your own Claude Code sign-in.
• Gemini, local models and other providers are not available yet. They are planned.
• There is no spending limit per task or project. Watch token usage in Reports.