Skip to content

Available AI agents

At sdd init you install exactly one AI coding agent host. Switch later with sdd agents install --ai <host> --force.

Hosts

Init flagProductWhat gets installedHow you use it
--ai copilotGitHub Copilot.github/agents/*.agent.md, AGENTS.md, .sdd/protocol.mdCopilot Chat → agent sdd
--ai grokGrok Build.grok/rules/sdd.md, AGENTS.md, .sdd/protocol.mdgrok CLI / Grok Build
--ai claudeClaude Code.claude/agents/*.md, AGENTS.md, .sdd/protocol.mdclaude CLI
--ai ollamaOllama (local).ollama/sdd.md, AGENTS.md, .sdd/protocol.mdollama run <model> (local LLM)

Interactive init (no --ai) prompts you to pick one of these.

Ollama setup

  1. Install Ollama and ensure ollama is on PATH.
  2. Pull a model, e.g. ollama pull llama3.2 (or any chat model you prefer).
  3. Init SDD with Ollama:
bash
sdd init --here --ai ollama
# optional: export SDD_OLLAMA_MODEL=qwen2.5-coder   # default: llama3.2
sdd new "Fix empty list crash" -w hotfix -y

When sdd launches the agent it runs roughly:

bash
ollama run "$SDD_OLLAMA_MODEL" "<kickoff from handoff>"

Ollama does not auto-load multi-role agent trees like Claude. Project guidance lives in:

  • .ollama/sdd.md — thin SDD router brief
  • AGENTS.md — host pointer
  • .sdd/handoff.md + .sdd/active-context.md — live task

Set the model with SDD_OLLAMA_MODEL or OLLAMA_MODEL (default llama3.2).

Agent roles (stubs)

Role idPurposeCopilot / ClaudeGrok / Ollama
sddRouter from current stageYesSingle brief file
sdd-plannerSpecs onlyYesUse router + stage
sdd-implementerCode for active changeYesUse router + stage
sdd-reviewerCheck acceptanceYesUse router + stage

Real process rules: .sdd/protocol.md. Live task: .sdd/active-context.md / .sdd/handoff.md.

Process commands vs agent launch

Most process commands refresh handoff and launch the configured agent.
Do not launch: status, init, workflows, context, doctor.

Skip launch once: --no-agent or SDD_NO_AGENT=1.

Switch host later

bash
sdd agents install --ai ollama --force
sdd agents refresh