Providers
The AI tools agents run on: what they have in common, where they differ, and how to pick one.
A provider is the AI tool that actually does an agent's work. Bounc drives thirteen of them; which one sits behind an agent is decided by the machine the agent is bound to.
What is the same
For the person handing out work there is barely a difference: every provider produces a transcript in one format, reports token usage, and obeys a cancellation. So agents on different tools look the same on an issue and land in the same Monitoring.
Where they differ
- Model choice. Some ship a built-in list, some discover it themselves, some take the model from the agent's settings.
- Continuing a conversation. Some providers can return to the previous session, so handing work over again picks up where it left off; the rest start over each time.
- Usage accounting. One provider does not report the tokens it spent — its runs show in monitoring by time and task count, but not by cost.
The catalogue
| Provider | Model choice | Session continuation |
|---|---|---|
| Claude Code | Built-in list | Yes |
| Codex | On the agent: model and reasoning effort | Yes |
| GitHub Copilot | Whatever your Copilot subscription allows | Yes |
| Gemini | On the agent | Yes |
| Antigravity | On the agent | Yes, but usage is not reported |
| Cursor | Discovered from the tool | No |
| Hermes | Discovered when connecting | No |
| Kimi | On the agent | No |
| Kiro | On the agent | No |
| OpenCode | Discovered from its own configuration | No |
| OpenClaw | Tied to the registered OpenClaw agent | No |
| Openclaude (Docker) | Built-in list, Claude-compatible | Yes |
| Pi | Discovered from the tool | No |
The model, the reasoning level, and any extra arguments are set on the agent and applied to each of its runs. See Agents.
How to choose
- When in doubt, Claude Code. The fullest integration: model choice, session continuation, a detailed transcript.
- When you want your own model catalogue — OpenCode or Cursor: they take the list from their own configuration.
- When the economics matter — do not put cost-tracked processes on Antigravity: its runs reach monitoring without usage.
- When the work is long and multi-step — pick a provider with session continuation, or every repeat run starts from nothing and spends the tokens again.
Access to external systems
Access to outside services used to be configured as MCP on the agent itself. It is connectors now: you connect a service once, grant it to the agents that need it, and revoke it in one move. Connectors also cover your own MCP servers, local and remote.
How fully a provider uses MCP depends on the tool: it works most completely with Claude Code and the Claude-compatible Openclaude.
Next
- Agents — where the provider and model are chosen.
- Computers and machines — where a provider runs.
- Connectors — granting an agent access to external systems.
Computers and machines
Where an agent's turn runs: the server, your computer or «No computer». How to connect a computer, what the instance admin decides, and what to do when an agent takes no work.
CLI reference
The bounc command groups, authentication, global flags, machine-readable output, and exit codes.