agent-generator¶
Describe an AI team in plain English. Get a runnable project.
You write one sentence. agent-generator picks a framework, wires the agents, generates the code, validates it, and hands you a ZIP you can unpack and run.
What you can build¶
- CrewAI projects — Python + YAML, ready for
crewai run - LangGraph state machines — typed graphs with one node per task
- WatsonX Orchestrate YAML — import straight into Orchestrate
- CrewAI Flow pipelines — event-driven mini-crews
- ReAct loops — single-file reasoning + tool agents
Run on WatsonX out of the box, or switch to OpenAI with one flag.
60-second tour¶
pip install agent-generator
agent-generator "Build a research team with a researcher and a writer" \
-f crewai -o team/
cd team && crewai run
That's it. No prompt engineering, no wiring tools by hand, no copy-pasting boilerplate. The LLM is called once to produce a structured plan; the rest is template rendering, so the same prompt always yields the same code.
Why it's safe to ship¶
| Concern | How we handle it |
|---|---|
| Hallucinated code | Templates render from a validated ProjectSpec, not from raw LLM output. |
| Unsafe patterns | AST scanner blocks eval, exec, os.system, bare subprocess. |
| Drift between runs | Same prompt + same spec = byte-identical output. |
| Secrets in code | Generated .env.example only; real secrets live in the platform's vault. |
| Audit trail | Every generation is logged; the platform stores the spec and the diff. |
Two ways to use it¶
| CLI | Platform | |
|---|---|---|
| Install | pip install agent-generator |
make install from a clone |
| Output | Files to disk | Web · Desktop · Android |
| Best for | Scripts, CI, one-offs | Teams, enterprises, marketplaces |
The CLI is the original tool. The platform adds a FastAPI backend, a SPA, desktop installers (Tauri), an Android app (Capacitor), audit logs, a project marketplace, and Helm charts. Same generator under the hood.
Also: the Matrix Builder engine¶
The same package is the deterministic engine behind Matrix Builder — give AI coders a contract, not a prompt. It turns an idea into a controlled bundle (blueprint, locked standards, scoped tasks, per-coder prompts), then validates what your AI coder produced and records each accepted change as an immutable Matrix Commit.
mb— local-first CLI: the full git-for-AI loop on your machine, offline and deterministic —mb init → next → prompt → check → timeline. See mb — local-first CLI.AgentGeneratorSDK:parse_idea,generate_blueprint_candidates,generate_controlled_blueprint,generate_matrix_bundle,generate_coder_prompt_pack,validate_ai_coder_patch. See Public engine API.- HTTP API:
uvicorn agent_generator.http.app:app(/api/v1/...). See HTTP API. - Works with Claude Code, Codex, Cursor, GitPilot, IBM Bob, and any AI coder; publishes validated bundles to MatrixHub under the Ruslan Magana Definitions.
Start here: Matrix engine quickstart.
Next: Install in 30 seconds · Usage recipes · Pick a framework · Matrix engine