> ## Documentation Index
> Fetch the complete documentation index at: https://docs.forge.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Coding Agents
> Discover coding agents and govern their model and MCP traffic through Forge routing.
Forge secures coding agents through their network traffic rather than by
instrumenting each agent. Supported requests are routed through the LLM Gateway
or MCP Gateway, where Forge applies identity, access, content, tool, and response
policies and records normalized sessions.
## Support
| Agent | Routed sessions | Inventory | Managed configuration |
| ------------- | --------------- | --------- | --------------------- |
| Claude Code | Supported | Supported | Supported |
| Claude Cowork | Supported | Supported | Supported |
| Codex | Supported | Supported | Supported |
| Cursor | Supported | Supported | Supported |
| OpenClaw | Supported | Supported | — |
| Rovo Dev | Supported | — | — |
Support is capability-specific. Inventory coverage does not imply that every
model API, MCP transport, or response format is available for content-aware
enforcement.
## Routing
Use one or both routing paths:
| Path | Use |
| ----------------- | ------------------------------------------------------------------------------------------------------------------------------------------------- |
| Network | Agentless routing through a supported firewall, SASE, or secure web gateway; preferred when enterprise AI traffic already traverses that platform |
| Forge for devices | Transparent endpoint routing for local or off-network use, with process and device attribution |
```mermaid theme={"system"}
flowchart LR
A["Coding agent"] --> B{"Routing point"}
B -->|Firewall, SASE, SWG| C["Forge rerouter"]
B -->|Forge for devices| D["Transparent proxy"]
C --> E["LLM Gateway"]
C --> F["MCP Gateway"]
D --> E
D --> F
E --> G["Model provider"]
F --> H["Approved MCP server"]
E --> I["Sessions, policies, violations"]
F --> I
```
Agentless network routing uses managed TLS inspection for the supported
content-aware destination set. Forge for devices provides the same gateway path
without modifying the coding agent or installing a Forge adapter inside it.
## Sessions
Forge reconstructs coding-agent sessions from routed model and MCP traffic. A
session preserves the user or workload identity, device or network source,
client and provider, native request identifiers, policy decisions, latency,
and upstream outcome.
| Activity | Available evidence |
| -------------- | ------------------------------------------------------------------------- |
| Prompt | Supported user or agent input before the model request |
| Model response | Supported model output before it returns to the client |
| MCP discovery | Server and tool definitions returned through the MCP Gateway |
| Tool call | Tool identity, structured arguments, access decision, and upstream result |
| Routing | Original destination, selected gateway route, provider, and bypass state |
Content visibility follows the selected privacy mode and the capability of the
protocol being routed. Forge records metadata-only traffic as metadata-only; it
does not present it as inspected content.
## Policies
The gateway evaluates the same Forge policy model used across other controlled
AI surfaces:
| Checkpoint | Evaluation |
| ---------- | ------------------------------------------------------------------------- |
| Prompt | Inspect, redact, guide, review, or block supported model input |
| Pre-tool | Authorize, redact, block, or require approval before an MCP tool executes |
| Post-tool | Inspect, redact, filter, or block a supported tool result |
| Response | Inspect or transform supported model output before returning it |
Access policies govern which agents, models, providers, destinations, MCP
servers, and tools an identity may use. Content policies govern the data and
actions inside an allowed interaction.
## Inventory
Device, repository, SaaS, and cloud collectors can discover coding-agent
configuration independently of routed activity:
| Artifact | Examples |
| --------------- | -------------------------------------------------------------- |
| Runtime | Installed application or CLI, version, and source path |
| Configuration | User, project, workspace, and system settings |
| Agent controls | Native permissions, commands, subagents, and existing hooks |
| AI dependencies | Models, providers, backends, MCP servers, and connectors |
| Extensions | Plugins, skills, instruction files, and workspace context |
| Change evidence | Configuration generations, content digests, and detected drift |
Hooks in Inventory are customer or product artifacts Forge discovered. Forge
does not install them to secure the agent.
## Configuration
Forge can manage supported native configuration for Claude Code, Claude Cowork,
Codex, and Cursor. Targets can be scoped by organization, team, device, system,
user, project, or session.
Managed settings can include tool permissions, filesystem scope, model and
provider selection, MCP connections, extensions, instruction files, telemetry,
and update behavior. Device Agent materializes the intended configuration,
records its generation, and reports drift or apply errors.
## Deployment
For agentless coverage, configure a supported Network integration and enable
content-aware routing for the coding-agent destinations in scope.
For endpoint coverage, deploy [Forge for devices](/integrations/device-agent)
directly, through MDM, CrowdStrike RTR, or SentinelOne. Confirm enrollment,
transparent-proxy health, managed CA trust, and a routed test request.
Fleet operations include installing or updating Device Agent, validating and
repairing the routing stack, refreshing credentials, removing managed
coverage, and offboarding a device.