CLIENT CONNECTIONS
Connect AI Clients to Antigravity Manager
Connect your AI client with three settings: the Manager gateway address, its API key, and a model your account can use. Find your client below for the right provider and URL.
Complete the first API request, then reuse its key, port and working model ID here. Keep a copy of your client’s current settings before adding the new connection.
Find the setup for your client
Choose your application and variant, then follow its setup link. These addresses use port 8045 with Manager and the client on the same computer; replace the host and port for your installation.
13 client variants
On a small screen, scroll the table sideways to see the address and setup.
| Client | Connection settings | How to connect |
|---|---|---|
| ClineEditor extension | OpenAI Compatiblehttp://127.0.0.1:8045/v1 | Set the provider in Cline settings. |
| Kilo CodeEditor extension | Gemini with a custom endpointhttp://127.0.0.1:8045 | Use the extension setup below. |
| Kilo DesktopDesktop app | Custom AI Provider → OpenAI Compatiblehttp://127.0.0.1:8045/v1 | Add a provider in AI settings. |
| Cherry StudioDesktop app | Custom provider → OpenAIhttp://127.0.0.1:8045 | Check the final request path in the form. |
| AlmaDesktop app | Custom providerhttp://127.0.0.1:8045/v1 | Add the model ID after saving. |
| Hermes AgentNous Research agent | Custom endpointhttp://127.0.0.1:8045/v1 | Run hermes model to configure it. |
| CursorBuilt-in chat / agent | Requests pass through Cursor servers A laptop’s loopback address is unreachable. | For a local gateway, use Cline in the editor. |
| CodeBuddyIDE | OpenAI; full request URLhttp://127.0.0.1:8045/v1/chat/completions | Add a model to models.json. |
| JeikCodeTerminal / TUI | Anthropic; TOML basehttp://127.0.0.1:8045/v1 | Sync in Manager or edit config.toml. |
| Claude CodeCLI | Anthropichttp://127.0.0.1:8045 | Open the Claude Code setup guide. |
| OpenCodeManager sync configuration | Anthropic adapterhttp://127.0.0.1:8045/v1 | Open the OpenCode setup guide. |
| CodexCLI | OpenAI Responseshttp://127.0.0.1:8045/v1 | Use Manager’s Codex sync card. |
| DeepSeek HarnessDSH · Manager 5.0.1-beta.2 | OpenAI Chathttp://127.0.0.1:8045/v1 | Sync the desktop and web profiles. |
No matching client. Clear the search to see every variant.
Which API address should you enter?
Use the address expected by the client’s provider. Manager accepts several API formats, and each client adds its own part of the request path.
OpenAI Chat or Responses
A base of http://127.0.0.1:8045/v1 lets the client add /chat/completions or /responses.
Anthropic Messages
Claude Code adds /v1/messages to the root address. The JeikCode TOML and OpenCode sync configurations already include /v1 in their base.
Gemini
Use http://127.0.0.1:8045 when the client adds the Gemini route beginning with /v1beta/models/.
Two useful exceptions: Cherry Studio’s API Host can add the full /v1/chat/completions suffix. CodeBuddy IDE’s url field takes the complete request URL. Follow their sections below.
Connecting from another machine? Enter the address reachable from the process running the client. On a laptop, phone or container, 127.0.0.1 refers to that environment itself. Use the remote-access guide to configure the gateway address and access controls.
Cline: add an OpenAI Compatible provider
In the Cline editor extension, open settings and choose OpenAI Compatible as the API provider.
- Set Base URL to
http://127.0.0.1:8045/v1. - Enter your Manager gateway API key and the model ID from your successful first request.
- Save the settings, select that model, and open a new task. If Plan and Act have separate provider settings, configure each mode you use.
Send a short message, then try the file-reading check. The same steps apply to Cline running as an extension inside Cursor. See Cline’s provider settings for the available model options.
Kilo Code: choose the extension or Desktop setup
The Kilo Code extension uses the Gemini setup below; Kilo Desktop has a separate custom-provider form.
Editor extension
In provider settings, choose Gemini and set its custom base URL to http://127.0.0.1:8045. Enter the Manager key and select your working model. This is the route recommended in Manager’s Kilo Code instructions.
If your extension has no custom Gemini address field, check its provider options before entering the key. If an OpenAI setup returns a 404 at /v1/chat/completions/responses, switch to the Gemini configuration or correct the selected adapter’s base URL.
Kilo Desktop
Open the sidebar gear, then AI Providers → Add AI Provider → Custom AI Provider. Choose OpenAI Compatible, enter http://127.0.0.1:8045/v1, the gateway key and your model ID. Save it and select the model for a new conversation. The Desktop AI settings also offer separate Responses and Anthropic formats; use the format required by your workflow.
Cherry Studio and Alma: add a chat provider
Create a provider named “Antigravity Manager” and select its model in a new conversation. Both apps let you keep your existing providers.
Cherry Studio
- Open Settings → Model Provider and add a custom provider. Choose the OpenAI connection format.
- Enter the gateway key and
http://127.0.0.1:8045as the API Host or OpenAI endpoint. Check the path preview: the final chat URL should behttp://127.0.0.1:8045/v1/chat/completions. Adjust the base if your version appends a different suffix. - Save, use Sync models to add your model, and enable the provider. Select the model in a new chat and send a short message.
Cherry Studio explains the fields in custom-provider setup and the suffix rules in API Host settings.
Alma
Open Settings → Providers → Add Provider, select Custom, and enter a name, http://127.0.0.1:8045/v1 and your gateway key. Save, then use Add Model to enter the exact model ID. Choose it in the conversation. See Alma’s custom-provider instructions.
Hermes Agent: configure a custom endpoint
In Nous Research’s Hermes Agent, run hermes model and choose Custom endpoint (self-hosted / VLLM / etc.).
- Enter
http://127.0.0.1:8045/v1as the API base URL. - Enter the gateway API key and your working model ID when prompted.
- Start a new Hermes session and send a short message. After changing providers, end the previous session before testing again.
This selects the OpenAI Chat Completions connection described in Hermes’ provider guide. If you use Hermes on a server, set an address reachable from that server.
If a Gemini request fails with safety_settings[4], save the exact error and model ID, then retry a short text prompt with optional request customizations removed. Compare the result through the direct API request to narrow down whether the failure starts in Hermes or the gateway. Include that comparison in a redacted bug report.
Cursor and CodeBuddy: connect from the right place
CodeBuddy IDE accepts a complete custom API URL. Cursor’s built-in chat sends requests through Cursor’s servers, which affects access to a local gateway.
Cursor
For a gateway running only on your computer, use Cline inside Cursor. Cursor’s own-key documentation explains that its built-in requests pass through its backend; that backend cannot reach your computer at 127.0.0.1. A custom endpoint for the built-in client needs both support in your installed version and an address reachable from the requesting server. Keep gateway authentication enabled for remote access.
CodeBuddy IDE
Add a model entry to ~/.codebuddy/models.json using the IDE’s configuration format. Replace MODEL_ID and GATEWAY_API_KEY with your values; merge the entry into an existing models array.
{
"models": [
{
"id": "MODEL_ID",
"name": "Manager — my model",
"vendor": "antigravity-manager",
"apiKey": "GATEWAY_API_KEY",
"url": "http://127.0.0.1:8045/v1/chat/completions"
}
]
}Reopen the IDE and select the model. If it is missing, check availableModels: an active list must include your model ID. Also check .codebuddy/models.json in the project, which takes priority over matching user settings. Store the key in your user configuration outside version control. Enable supportsToolCall or supportsImages only for features the selected model supports.
JeikCode: sync or configure the connection manually
Manager 5.0.0 can configure JeikCode’s Anthropic connection from its agent-tools panel. Install JeikCode on the machine where you will run it, then choose one of these methods.
Use Manager’s sync action
- Save a dated copy of your JeikCode configuration. Sync replaces Manager model entries and can consolidate accounts that use the same gateway URL.
- Start API Proxy and open One-Click Agent Tools Configuration. On the JeikCode card, select your model and click Sync Config Now.
- Open the generated configuration and check the address, port and selected model. Run
jeikcodeto start the terminal interface;/webuiopens its web interface from that session.
The file is ~/.jeikcode/config.toml, or config.toml inside JEIKCODE_HOME when that variable is set. On Windows, the default folder is .jeikcode under your user profile. For a remote Manager, read where sync writes files before using it.
Set up config.toml manually
Use this single-model configuration from Manager’s JeikCode integration. Replace every MODEL_ID with the same working model ID and replace GATEWAY_API_KEY locally. Merge the entries into your existing file, keeping unrelated settings.
default_model = "MODEL_ID"
default_provider = "MODEL_ID"
[provider_accounts.antigravity-manager]
provider = "anthropic"
base_url = "http://127.0.0.1:8045/v1"
api_key = "GATEWAY_API_KEY"
[models."MODEL_ID"]
account = "antigravity-manager"
model = "MODEL_ID"Keep both default_model and default_provider set to the model ID: this is the format Manager’s JeikCode integration uses. The TOML base includes /v1.
Try a temporary Bash session
On macOS or Linux, run bash if your current shell is different, then paste this command. It asks for the key with hidden input and applies the connection settings to this JeikCode process. Change port 8045 first if needed.
(
read -r -s -p "Gateway API key: " AG_CLIENT_KEY || exit 1
printf '\n'
read -r -p "Available model ID: " AG_CLIENT_MODEL || exit 1
if [[ -z "$AG_CLIENT_KEY" || -z "$AG_CLIENT_MODEL" ]]; then
printf 'Enter both a key and a model ID.\n' >&2
exit 1
fi
ANTHROPIC_BASE_URL="http://127.0.0.1:8045" \
ANTHROPIC_API_KEY="$AG_CLIENT_KEY" \
jeikcode --model "$AG_CLIENT_MODEL"
)The environment-variable method uses the gateway root. The subshell discards the entered values when JeikCode exits. For desktop or web launches, check the configuration loaded by that process and the address it can reach.
BETA FEATURE
DeepSeek Harness: sync the desktop and web profiles
Manager 5.0.1-beta.2 includes DSH sync for OpenAI Chat Completions. Use a Manager build with this feature; stable 5.0.0 users can configure their client manually or choose the beta release.
Back up your DSH settings and credentials. On the machine and OS user running DSH, start the gateway and run the beta Manager command with your model ID:
antigravity-tools sync dsh --model MODEL_IDThe command reads Manager’s configured key and port, then writes a loopback base ending in /v1 with api: openai-completions. Check these files before starting DSH:
- CLI and desktop:
~/.dsh/profiles/desktop/cordis.patch.yml - Web interface, started with
dsh web:~/.dsh/profiles/web/cordis.patch.yml - Credentials:
~/.dsh/.credentials.yaml
The beta.2 sync action writes under the OS home directory, even if you set DSH_HOME. For a custom DSH home, update the files DSH actually loads. For a remote gateway, replace the loopback address in both profiles with its reachable address. See the beta CLI’s sync targets for the command and paths.
Claude Code, OpenCode and the machine sync writes to
Use the Claude Code connection guide or the OpenCode connection guide for their full setup and restore steps. For Codex CLI, Manager’s sync card writes a Responses provider with a base ending in /v1.
Run sync as the OS user who runs the client. In a container, keep the configuration and credentials in persistent storage. After changing a CLI configuration, restart the client and select its new provider. For a separately installed desktop app, check that app’s provider settings and the file it reads.
CC Switch is optional. It can manage client profiles, prompts, Skills and MCP settings. Vibe Kanban manages coding tasks and working copies. A direct client connection needs only the gateway settings described here.
Test the connection and restore your previous settings
Test a short conversation first, then the feature you intend to use. Keep Manager’s request logs open so you can match each test to the selected route and model.
- Chat: start a new conversation and ask for a short greeting. Read the answer and confirm that Manager received the request.
- Tools: in a small test folder, create a text file containing a unique marker. Ask the agent to read the file with its file tool and return the marker. Check the tool call, result and final answer. Chat-only clients can skip this step.
- Follow-up and streaming: ask a second question about the answer or marker. Let the response finish. For a client you use across restarts, close and reopen it, then repeat a short request.
Connection refused: check the running service, host and port. 401 or 403: check the gateway key and its access rules. 404: inspect the final URL for repeated path segments; if the route is correct, read the error for an unavailable model. Text works but tools or streaming fail: follow the tool and streaming troubleshooting guide.
For a rejected parameter such as safety_settings, keep the exact field name and retry a minimal request before changing model options. Use the diagnostics guide to report the Manager/client versions, selected provider, redacted endpoint, error and reproduction steps.
To undo the setup: close the client, restore your saved configuration or select the previous provider, and reopen it. If using Manager’s Restore Backup, check the backup date first: an existing automatic backup may predate your latest edits. Close any temporary test session and stop the proxy when you have finished using it.
Choose your next step
For connection failures, start with host, port and routing checks. Once chat works, use model selection and routing to choose the model for each task.
- Connect Claude Code — environment settings, sync and recovery.
- Connect OpenCode — provider configuration and model selection.
- API reference — endpoints and request formats.
- Manager 5.0.0 documentation — application setup and client integrations.