Agents
Set up coding agents, cloud providers, and models running on your computer.
Loading page content...
Set up coding agents, cloud providers, and models running on your computer.
Loading page content...
Choose the AI tools for your Glowbom Desktop workflow. Install a coding agent and connect its provider, or use a model running on your computer. You only need one connection to start. Each connection keeps its own models, permissions, account access, and usage limits. The coding-agent setup also applies to OSS.
Both clients use the open backend on your computer. Chat, Draw, Studio, and Project Book screens below belong to the full Desktop client; OSS has the simpler Agent + Buzz workspace.
| Connection | Use in Glowbom | Setup |
|---|---|---|
| OpenCode | Chat, Draw, Project Book, and Build with suitable connected models. | Glowbom OSS setup and OpenCode providers. |
| Codex App Server | Chat, Draw, Project Book, and Build with supported models. | Install Codex CLI and sign in below. Glowbom starts the server. |
| Cursor | Build in your project folder. | Install and sign in to the CLI below. |
| Claude Code | Build in your project folder. | Claude Code setup. |
| ACP agents | Build through a manually saved connection. | Configure the agent below. |
| Local models | Chat or Build, depending on the model and connection. | Local AI setup with Ollama and OpenCode. |
Chat discusses an idea. Draw makes a browser prototype. Build lets an agent change project files and run commands. Selecting Cursor, Claude Code, or ACP switches the composer to Build.
Local AI is another connection choice. The local AI setup guide covers Ollama models, local voice, and dictation. Start there if you want a model running on your computer, then select the connected model in Glowbom. Desktop has guided setup screens; OSS can use suitable coding models through OpenCode's configuration. Model capability and memory requirements vary.
| Agent | Installation | Authentication |
|---|---|---|
| OpenCode | OpenCode installation | opencode auth login |
| Codex | Codex CLI installation | codex login, or Glowbom's Sign in with ChatGPT button. |
| Claude Code | Claude Code setup | claude auth login |
| Cursor | Cursor CLI installation | See the command-name and path instructions below. |
Run opencode auth login, choose a provider, and finish its sign-in or API-key
setup. Glowbom uses OpenCode's connected providers and models. Available login
methods depend on the provider and OpenCode version. Follow the
provider guide for your account.
The launcher checks the OpenCode version and selects its matching adapter.
Install Codex CLI, then run codex login and complete the browser sign-in.
You can also select Codex in Glowbom and use Sign in with ChatGPT.
Check terminal authentication with codex login status; see
Codex authentication for account
options and headless setup.
Glowbom starts Codex App Server over its local stdio connection. There is no separate server command to run or port to configure. After connecting, choose a model reported by Codex. Model availability and usage depend on your account.
Install the CLI, run claude auth login, then check claude auth status.
These commands are documented in the
Claude Code CLI reference.
Select Claude Code in Glowbom's Build controls. This connection runs the
Claude Code CLI in your project folder; it does not add general Chat or Draw
models. If you installed the CLI while Glowbom was open, refresh the tool checks.
Follow Cursor's CLI installation guide. On macOS, Linux, or Windows through WSL:
curl https://cursor.com/install -fsS | bashWith the cursor-agent command Glowbom detects, sign in and check the models:
cursor-agent login
cursor-agent status
cursor-agent modelsFinish the browser sign-in. Installing the Cursor editor alone is not enough.
Ready in Tools means Glowbom found the executable; it does not confirm
account access. Glowbom only lists Cursor models when the CLI's models
command succeeds.
After signing in, allow 45 seconds for the model cache to expire. In OSS, open Settings → Refresh. In Desktop, choose Refresh models; Cursor → Add models exposes more models reported by your account. If the entire Cursor group is missing, first check CLI login and model discovery.
Cursor's current authentication
and command reference use
agent. If your installation provides only that command, run
agent login, agent status, and agent models. For the OSS terminal
workflow, find its path with command -v agent, set
GLOWBOM_CURSOR_BIN to that full path, then launch glowbom start from the
same terminal. This override applies to that OSS launch.
Agent Client Protocol (ACP) connects coding agents to other tools. Glowbom can save up to three installed agents, including custom agents.
These are manual setup examples, not built-in presets. Follow each agent's linked installation guide and test your installed version.
| Agent | Executable | Arguments | Before connecting |
|---|---|---|---|
| Cline | cline | --acp | Install with npm install -g cline, then run cline auth. |
| Goose | goose | acp | Run goose configure; enable its Developer extension. |
| Grok Build | grok | agent and stdio, one per line | Run grok login; follow the testing steps below. |
| Hermes Agent | hermes | acp | Configure with hermes model; check hermes acp --check. |
| Kilo | kilo | acp | Complete its provider and account setup. |
| OpenClaw | openclaw | acp | Configure and start its Gateway. |
For Cline, choose its model in Glowbom's connection settings; do not assume its terminal selection is the ACP default. Test again after changing the agent's provider or account.
For Goose, install its CLI
and run goose configure to save your provider and model. Enable its Developer
extension for file edits and commands. Glowbom uses Goose's saved setup; provider
API keys and GOOSE_* variables from Glowbom's environment are not passed through.
On macOS, check for a Keychain prompt if connecting stalls. Use Test connection,
then try a small Build to check your Goose version and provider.
OpenClaw's ACP command bridges to a running Gateway. Complete OpenClaw's setup, start the Gateway, and check that its agent can work in your project folder. Model selection stays in the Gateway. Tool details and approval relay are limited by the bridge. Keep Gateway credentials in OpenClaw's saved setup, outside Glowbom's argument fields.
These commands expose ACP; a successful connection test is the first check. Try a small build before relying on an agent for your project.
Use a full executable path if Glowbom cannot find the program. Keep shell quotes, commands, and credentials out of these fields. Save does not start the agent. Test connection does not verify a build or every agent feature. An unavailable or rejected saved model stops the build with an error.
With an ACP-only setup, glowbom doctor can still report a missing coding agent:
that check currently recognizes the four built-in CLI integrations. This does
not prevent glowbom start. Use Test connection and a small Build to check
your ACP agent, while resolving any missing Go, Bun, or checkout requirements.
Install Grok Build, then run grok login
and finish the browser sign-in. In an empty ACP slot, use Grok Build as the
name and grok as the executable. Enter these arguments on separate lines:
agent
stdioThis starts the documented Grok ACP connection.
Glowbom requests headless authentication using the saved login when the agent
advertises its cached_token authentication method. An expired or missing
login requires signing in again in Terminal. Provider API-key environment
variables are not forwarded, and keys do not belong in the argument fields.
Choose Test connection, select a reported model or Use agent default, and save. In a disposable project, ask Grok Build to create a short text file. Check the complete final response, the file on disk, permission requests, and Stop before using it on important work. This source adds the cached-login handshake; real Grok Build acceptance is still pending. Restart a backend running older source to use it; installed Desktop packages need an update containing this change.
Install Hermes with the optional ACP dependency from its ACP setup guide. Configure the provider and model in Terminal:
hermes model
hermes acp --checkAdd a connection named Hermes, executable hermes, and argument acp.
Choose Test connection, select a reported model or Use agent default,
then save. Hermes uses its own saved provider configuration and tools. If the
check reports a missing ACP dependency, complete the optional installation
step in the linked guide before trying again.
Try a small Build in a disposable folder and check the resulting files, permissions, and Stop. These are setup instructions, not confirmation of a successful run with your Hermes version and provider.
ACP builds use Glowbom's progress, permission, Stop, and history controls. Allow for this build approves a named tool when the agent offers that choice. Allow all for this build starts off and approves supported tool requests until the build ends. Questions and paid-media reviews still need an answer. Cursor already runs tools automatically under its CLI deny rules.
ACP agents must provide their own file and command tools. Image input and conversation restoration depend on the agent's capabilities. ACP does not add Chat, Draw, or Project Book model support. Compatible builds can also be started from the companion preview. Setup and credentials stay on Desktop; messages and media go to the chosen provider through it.