Problem
Hand-authoring an AI agent package means getting a lot of interlocking pieces right at once: agent.yaml, SOUL.md, RULES.md, skill definitions, MCP-compatible tool schemas, workflows, knowledge docs — each with its own shape, and no single place to see how a change to one ripples into the rest. Tool schemas in particular have a specific, easy-to-miss requirement (input_schema, not parameters) that a hand-written file will quietly get wrong.
Approach
GitAgent Workbench is a guided wizard — Identity → Capabilities → Model → Compliance → Structure → Review — that builds a central AgentWorkspace state as you go, validated at each step with Zod schemas. The structure type you choose (minimal, standard, full, or a domain-specific preset like data-analyst or web-scraper) determines which of a 12-step generation pipeline actually runs, from manifest and identity generation through skills, tools, and — for full structures — multi-step workflows and example outputs.
Generation calls out to one of seven AI provider integrations (Anthropic, OpenAI, Google, Mistral, Groq, Ollama, OpenRouter) with retry logic and a local-synthesis fallback if a provider call fails. Once generation completes, a serializer packages everything into a downloadable ZIP: stripping null values, adding YAML frontmatter to skill files, and normalizing tool schemas to the format gitagent validate expects.
An optional sidecar, the local context backend, adds hybrid RAG over your existing agent configs — chunking with tiktoken, storing in SQLite with FTS5 (BM25) and sqlite-vec, and merging keyword and vector results with Reciprocal Rank Fusion — so generation can be grounded in context you already have rather than starting from a blank page.
Status
Alpha. The wizard, the 12-step generation pipeline, and ZIP export are implemented end-to-end. The local context backend’s embedding generation is still TODO in the project’s own docs — retrieval falls back to keyword search until that lands.