Skip to content

Configuring agents

After creating an agent from a preset, you can customise its model, permissions, tool access and behaviour. Right-click any agent in the sidebar and select Reconfigure to open the configuration modal.

Model and Provider Selection

By default, agents use the model set in your global settings. You can override both the provider and model per agent.

The model dropdown shows all models from your configured providers, grouped by provider family. Each entry displays capability indicators:

IndicatorMeaning
Context windowMaximum token capacity (e.g. 200K, 1M)
Tool supportWhether the model can use tools (Read, Edit, Bash, etc.)
Vision supportWhether the model accepts image input

Ironspire ships with static model registries for seven provider families (Claude, OpenAI, Google Gemini, Mistral, DeepSeek, xAI Grok, AWS Bedrock). Ollama models are discovered at runtime. Custom OpenAI-compatible endpoints can be added in Settings > Providers.

See the model registry for the full list of supported models and their capabilities.

Effort Level

Controls how deeply the agent reasons before responding:

LevelDescription
LowFast, minimal thinking. Best for simple tasks.
MediumBalanced reasoning and speed.
HighDeep reasoning. Best for complex tasks. (Default)

Lower effort is faster and cheaper. Use Low for straightforward tasks like formatting or simple lookups, and High for architecture decisions or debugging.

Permission Mode

Controls whether the agent asks before using risky tools:

  • Normal (Safe): the agent prompts for permission on risky actions like file writes and shell commands. This is the default.
  • Bypass Permissions (Caution): the agent uses tools without asking. Use this only for trusted, well-understood workflows.

Bypass Permissions lets the agent modify files, run commands and make changes without confirmation. Make sure you understand what the agent will do before enabling this.

Tool Access

You can enable or disable individual tools per agent:

ToolWhat it does
ReadRead file contents
EditModify existing files
WriteCreate new files
GlobSearch for files by pattern
GrepSearch file contents
BashRun shell commands
WebSearchSearch the web
WebFetchFetch content from URLs

Disabling a tool removes it from the agent entirely. This is useful for restricting agents to read-only access or preventing shell commands.

Docker Sandboxing

You can run individual agents inside a Docker container for additional isolation. Toggle Docker sandbox in the agent configuration modal to enable it.

When sandboxing is active, the agent's file system access and shell commands run inside a container rather than directly on your host machine. Docker Desktop must be installed and running.

See the Docker sandboxing guide for setup instructions and configuration options.

Compaction Override

If auto-compaction is enabled globally, you can override the settings for a specific agent:

  • Context threshold (60-90%): when to trigger compaction
  • Preserved turns (2-6): how many recent exchanges to keep

This lets you give a long-running agent more aggressive compaction while keeping short-task agents at default settings.

Permission and tool changes take effect after restarting the agent session.

Next steps