Configuration Reference
OpenFlows has no public environment-variable reference — that is internal detail. What matters to you are the knobs you actually set, grouped here by concern. Almost everything is changeable from the control panel without a restart. Platform-level items (which models are available, centrally managed tools) are configured in the Coder dashboard.
Team configuration
Each agent role — FORGE, SENTINEL, VESSEL, and LORE — is configured from the control panel. The same settings apply regardless of role:
| Setting | What it controls |
|---|---|
| Worker count | How many agents a role may run at once — its ceiling on concurrency. |
| Model | Which model the role uses. |
| Plan mode | Whether the role must get approval before acting. |
| Skills | Which packaged capabilities the role can use. |
| External tools | Which outside tools the role can call. |
A guardrail on changes
Fleet steering
| Setting | Effect |
|---|---|
| Pause | Let in-flight work finish; don’t pick up new work. |
| Drain | Stop taking on new tickets. |
| Target | Focus on a specific repository, issue, or label. |
| Resume / normal | Return to ordinary operation. |
Steering changes how the fleet picks up and processes work without requiring a restart — you can pause, drain, or refocus the team on the fly.
Platform-level settings
These live in the Coder dashboard rather than in OpenFlows itself, because they govern the shared environment all teams run on:
| Setting | What it controls |
|---|---|
| Models available | The catalog of models roles can reference. |
| Central tools | Team-wide tools with allow and deny lists and availability. |
| Identity / single sign-on | Who can reach the environment and its workspaces. |
| Audit | The platform audit log. |
Notifications
| Channel | What you set |
|---|---|
| Slack | A webhook URL. |
| Discord | A webhook URL. |
| Credentials and sending/destination phone numbers. |
Each channel is configured independently, and any combination can be enabled. See the Notifications Reference for how they behave.
Environment & topology
These are decided at install time and describe how your deployment is put together:
| Piece | What it provides |
|---|---|
| Coder server location | Where the control plane is. |
| Coordination store | The shared record of truth that tracks all team state. |
| Database | Coder’s own data store. |
| Repository binding | The repository a tenant works on. |
Core behaviors you generally don’t change
Some things are deliberate design decisions rather than toggles:
- The coordination store is the single source of truth — only the controller and the worker surface may write to it.
- The planning gate is structural — it can’t be turned off by a setting.
- “When in doubt, don’t approve” — the reviewer’s bias toward blocking over guessing.
- Recovery is bounded, then escalated. These are product invariants, not preferences.
Related
- Limits & Defaults — the meaningful numbers behind these knobs.
- Roles Reference — who the team configuration applies to.
- Notifications Reference — how the notification channels behave.