Create coding agent workspace
Add your code location first, then tune agent, tools, and sandbox access in the following steps.
Workspace
Local folder or Git remote — we detect it, analyze the tree, and name the workspace.
Path, HTTPS/SSH URL, or GitHub owner/repo. Use ./ for local folders that look like a/b.
Description (optional)
Saved project (optional)
Load defaults from a project
Path, skills, MCP, vault, knowledges, and access presets — same cards as on Projects.
Agent & model
Pick a coding agent, then choose its default model — same rhythm as onboarding.
Choose your coding agent
Pick one runtime for kdn in this workspace. API keys and providers are configured in Settings; the list below shows models that match the selected agent.
Select a Semantic Router. The agent will call http://localhost:{port}/v1/chat/completions using model: "auto" — the router picks the best backend for each request automatically.
Model for this workspace
Choose the default model OpenCode will use here. Names match the Models catalog. Disabled rows cannot be selected; the table is filtered to models that fit the agent you picked above (OpenCode: all sources; Claude Code: Anthropic cloud; Codex: OpenAI & Composer).
Cloud · LLM providers
| Status | Name | Size | Runtime | Use |
|---|---|---|---|---|
|
composer-2-fast
Cloud · enabled in catalog
|
— | Composer | ||
|
gpt-5.3-codex
Cloud · enabled in catalog
|
— | OpenAI | ||
|
gpt-5.4-medium
Cloud · enabled in catalog
|
— | OpenAI | ||
|
claude-4.6-sonnet-medium
Cloud · enabled in catalog
|
— | Anthropic | ||
|
claude-4.6-opus-high
Cloud · enabled in catalog
|
— | Anthropic | ||
|
composer-2
Cloud · disabled in catalog
|
— | Composer | ||
|
gpt-5.3-codex-low
Cloud · disabled in catalog
|
— | OpenAI |
In-house · OpenShift AI
| Status | Name | Size | Age | Use |
|---|---|---|---|---|
|
ibm-granite-3.3-8b-instruct
OpenShift AI · Apache-2.0
VRAM est. 8.2 GiB
|
4.8 GB | 2 weeks | ||
|
mistral-small-internal
OpenShift AI
VRAM est. 11 GiB
|
6.1 GB | 1 month | ||
|
llama-3.1-70b-instruct
OpenShift AI · pending approval
RAM usage: N/A
|
— | — |
Local · Ollama & Ramalama
| Status | Name | Size | Runtime | Use |
|---|---|---|---|---|
|
qwen3-code
GGUF · loaded in memory · OpenCode default
RAM usage: 5.8 GiB
|
5.4 GB | Ollama | ||
|
llama3.2:3b
GGUF · loaded in memory
RAM usage: 3.93 GiB
|
2.0 GB | Ollama | ||
|
mistral:latest
GGUF · loaded in memory
RAM usage: 4.4 GiB
|
4.1 GB | Ollama | ||
|
qwen2.5:7b
Not loaded — Ramalama stopped
RAM usage: N/A
|
4.7 GB | Ramalama |
Tools & secrets
Every skill, MCP server, vault credential, and listed knowledge base is included by default so you can go faster. Expand Customize when you want a smaller toolset.
Customize skills, MCP servers, vault, and knowledges
Skills
6 skillsMCP Servers
6 serversSecret Vault
6 credentialsKnowledges
4 knowledge basesOptional retrieval context for the agent. Manage environments in Knowledges.
Sandbox · Filesystem
Host file access for this workspace. Same presets as first-run setup — click a row or use Use.
File system
Same layout as Models — pick one scope for agent read/write on the host.
| Status | Name | Access | Notes | Use |
|---|---|---|---|---|
|
No host filesystem access
The agent cannot read or write files on your host. Use for API-only or fully remote workflows.
Recommended
|
None | Strict isolation | ||
|
Home Directory
Agent can access your entire home directory (~/) and all subdirectories.
|
Home (~) | Local development | ||
|
Custom Paths
Specify exact directories the agent can access.
|
Listed paths | Set path below | ||
|
Full System Access
Agent can access the entire filesystem. Use with caution.
|
Full host | High privilege |
Allowed path for custom access. You can refine paths later in project or workspace settings.
Use an absolute path on the host. Tilde (~) is expanded when the workspace starts.
Sandbox · Networking
Outbound network for this workspace sandbox. Separate from filesystem access on the previous step.
Network policy
Same layout as Models — click a row or use Use.
| Status | Name | Egress | Notes | Use |
|---|---|---|---|---|
|
Deny All
No outbound HTTP/HTTPS from the sandbox.
|
None | Strict | ||
|
Developer Preset
Allow npm, PyPI, and similar registries — not arbitrary public hosts.
Recommended
|
Registries | Balanced default | ||
|
Agent mode
The agent requests each outbound access; you approve before traffic leaves the sandbox.
|
Per request | Human in the loop | ||
|
Unrestricted
Permit all outbound traffic. Best for trusted dev setups.
|
All hosts | Trusted setups |
Allowlists and more — Fine-grained host allowlists and static egress rules live in project or workspace settings when you need them.