Seven role-specific reasoning agents that turn vague intent into precise operating artifacts.
v1.1.0 — Specification Kernel + Founder · Design · Product · Engineering · Marketing · Trust/Risk · CIO Agents + multi-agent pipeline orchestration + intra-stage progress events
AgentSuite is a Python package and MCP server. It exposes role-specific agents (Founder, Design, Product, Engineering, Marketing, Trust/Risk, and CIO shipped) that take loose human intent and produce structured, reusable artifacts: brand systems, brief libraries, voice guides, prompt templates, engineering specs, IT strategies, and more.
The agents are reasoning agents, not content generators. Output is a reusable system, not a one-off asset.
Technical founders and developers who build inside AI-native IDEs — Claude Code, Cowork, Codex, Google Antigravity — and are frustrated that every new session starts from scratch. You know your way around a terminal. You have an API key. You want structured, reusable output, not a one-off generation you'll never find again.
The problem: AI IDEs are great at generating content but terrible at remembering it. Every session re-introduces context, re-drifts on voice, re-argues decisions that were settled last week.
What AgentSuite does: acts as the operating layer between your intent and your IDE. Seven role-specific agents each walk a five-stage pipeline (intake → extract → spec → execute → qa), persist artifacts to disk, and promote approved output into a _kernel/ that every subsequent run — and every downstream agent — consumes. Your brand system, your engineering decisions, your risk posture: written once, versioned, reused.
What makes it different: the system around generation, not the generation itself. Every artifact is on disk, QA-scored, and reusable across sessions. Provider-agnostic (Anthropic / OpenAI / Gemini / local Ollama). MCP-native — agents surface as tools in Claude Code, Cowork, Codex, Google Antigravity, or any MCP-compatible IDE without a separate server.
# Install with your preferred provider:
pip install "agentsuite[anthropic] @ git+https://github.com/scottconverse/AgentSuite.git" # Anthropic Claude
pip install "agentsuite[openai] @ git+https://github.com/scottconverse/AgentSuite.git" # OpenAI GPT
pip install "agentsuite[gemini] @ git+https://github.com/scottconverse/AgentSuite.git" # Google Gemini
pip install "agentsuite[ollama] @ git+https://github.com/scottconverse/AgentSuite.git" # Local Ollama daemon
# Install everything:
pip install "agentsuite[all] @ git+https://github.com/scottconverse/AgentSuite.git"
# or, no install (MCP only):
uvx --from "agentsuite[mcp] @ git+https://github.com/scottconverse/AgentSuite.git" agentsuite-mcpAgentSuite is distributed from GitHub only — there is no PyPI publication.
Requirements: Python 3.11+. Either an API key (ANTHROPIC_API_KEY, OPENAI_API_KEY, or GEMINI_API_KEY / GOOGLE_API_KEY) OR a local Ollama daemon running at localhost:11434.
export ANTHROPIC_API_KEY=sk-ant-...
agentsuite founder run \
--business-goal "Launch PatentForgeLocal v1" \
--project-slug pfl \
--inputs-dir ./examples/patentforgelocalThe agent walks five stages, writes 26 artifacts under .agentsuite/runs/<run-id>/, and pauses at the approval gate. Review brand-system.md + qa_report.md, then:
agentsuite founder approve --run-id <run-id> --approver you --project-slug pflApproved artifacts get promoted to .agentsuite/_kernel/pfl/ for downstream agents.
Run AgentSuite entirely offline against a local Gemma 4 model. Zero cost, no API keys required.
# Install Ollama (https://ollama.ai) if not already installed, then pull a Gemma 4 model:
ollama pull gemma4:e4b # recommended: ~5 GB, balanced quality
# Optional alternatives:
# ollama pull gemma4:e2b # ~3 GB, laptop-class (faster, less capable)
# ollama pull gemma4:26b-moe # ~15 GB, high-end workstation (slowest, best quality)
# Make sure the daemon is running:
ollama serve &
# Then run AgentSuite as usual — the resolver auto-detects the daemon when no API keys are set:
agentsuite founder run \
--business-goal "Launch My Product v1" \
--project-slug my-product \
--inputs-dir ./my-brand-inputsYou can also pin Ollama explicitly via AGENTSUITE_LLM_PROVIDER=ollama regardless of which API keys you have set.
Add to ~/.codex/mcp.toml:
[servers.agentsuite]
command = "uvx"
args = ["agentsuite-mcp"]
[servers.agentsuite.env]
AGENTSUITE_ENABLED_AGENTS = "founder,design,product,engineering,marketing,trust-risk,cio"Restart Codex. Tools agentsuite_founder_run, agentsuite_founder_approve, agentsuite_founder_get_status, agentsuite_founder_list_runs, agentsuite_founder_resume, plus the cross-agent agentsuite_list_agents, agentsuite_kernel_artifacts, agentsuite_cost_report are now callable.
Add to project-root .mcp.json:
{
"mcpServers": {
"agentsuite": {
"command": "uvx",
"args": ["agentsuite-mcp"],
"env": {"AGENTSUITE_ENABLED_AGENTS": "founder,design,product,engineering,marketing,trust-risk,cio"}
}
}
}Restart the harness.
AgentSuite is MCP-compatible with Google Antigravity. Refer to the Antigravity documentation for the MCP server configuration format, then point it at uvx agentsuite-mcp with AGENTSUITE_ENABLED_AGENTS set to your desired agent list.
| Stage | Artifacts |
|---|---|
| 1 intake | inputs_manifest.json |
| 2 extract | extracted_context.json |
| 3 spec | brand-system.md, founder-voice-guide.md, product-positioning.md, audience-map.md, claims-and-proof-library.md, visual-style-guide.md, campaign-production-workflow.md, asset-qa-checklist.md, reusable-prompt-library.md, consistency_report.json |
| 4 execute | brief-template-library/ (11 brief templates) + export-manifest-template.json |
| 5 qa | qa_report.md, qa_scores.json |
| state | _state.json, _meta.json |
On agentsuite_founder_approve, the spec artifacts + brief-template-library are promoted to .agentsuite/_kernel/<project_slug>/ for use by downstream agents.
| Stage | Artifacts |
|---|---|
| 3 spec | architecture-decision-record.md, system-design.md, api-spec.md, data-model.md, security-review.md, deployment-plan.md, runbook.md, tech-debt-register.md, performance-requirements.md |
| 4 execute | brief-template-library/ (8 brief templates: sprint ticket, code review checklist, incident report, capacity plan, on-call handoff, release checklist, postmortem, vendor evaluation) |
| 5 qa | qa_report.md, qa_scores.json |
| state | _state.json, _meta.json |
Input: product_name, risk_domain, stakeholder_context (plus optional regulatory_context, threat_model_scope, compliance_frameworks).
| Stage | Artifacts |
|---|---|
| 3 spec | threat-model.md, risk-register.md, control-framework.md, incident-response-plan.md, compliance-matrix.md, vendor-risk-assessment.md, security-policy.md, audit-readiness-report.md, residual-risk-acceptance.md, consistency_report.json |
| 4 execute | brief-template-library/ (8 brief templates: breach-notification, executive-risk-summary, penetration-test-brief, remediation-tracker, risk-acceptance-form, security-awareness-brief, tabletop-exercise-scenario, vendor-security-questionnaire) + export-manifest-template.json |
| 5 qa | qa_report.md, qa_scores.json |
| state | _state.json, _meta.json |
Primary artifact: threat-model.md. On agentsuite_trust_risk_approve, all spec artifacts and brief templates are promoted to .agentsuite/_kernel/<project_slug>/.
Input: organization_name, strategic_priorities, it_maturity_level.
| Stage | Artifacts |
|---|---|
| 3 spec | it-strategy.md, technology-roadmap.md, vendor-portfolio.md, digital-transformation-plan.md, it-governance-framework.md, enterprise-architecture.md, budget-allocation-model.md, workforce-development-plan.md, it-risk-appetite-statement.md, consistency_report.json |
| 4 execute | brief-template-library/ (8 brief templates: board-technology-briefing, it-steering-committee-agenda, vendor-review-summary, project-portfolio-status, digital-initiative-proposal, it-investment-case, technology-modernization-pitch, quarterly-it-review) + export-manifest-template.json |
| 5 qa | qa_report.md, qa_scores.json |
| state | _state.json, _meta.json |
Primary artifact: it-strategy.md. On agentsuite_cio_approve, all spec artifacts and brief templates are promoted to .agentsuite/_kernel/<project_slug>/.
Use AgentSuite programmatically — no CLI required:
from pathlib import Path
from agentsuite import FounderAgent, resolve_provider
from agentsuite.agents.founder.input_schema import FounderAgentInput
from agentsuite.kernel.schema import Constraints
agent = FounderAgent(output_root=Path(".agentsuite"), llm=resolve_provider())
state = agent.run(
request=FounderAgentInput(
agent_name="founder",
role_domain="product",
user_request="My startup builds X for Y",
business_goal="Validate and launch",
project_slug="my-startup",
constraints=Constraints(),
),
run_id="my-run",
)
print(state.stage) # "approval"resolve_provider() auto-detects your API key from environment variables. See Configuration for the full list of env vars.
Run multiple agents back-to-back — each agent's output feeds into the next run's context via the shared _kernel/<project_slug>/ directory.
CLI — headless (auto-approve each step):
agentsuite pipeline run \
--agents founder,design,product \
--project-slug my-startup \
--business-goal "Launch a B2B invoicing tool for freelancers" \
--auto-approveCLI — with approval gates (default):
# Step 1: start the pipeline (pauses after founder)
agentsuite pipeline run \
--agents founder,design,product \
--project-slug my-startup \
--business-goal "Launch a B2B invoicing tool for freelancers"
# Prints: pipeline-id: pipeline-20260501T120000-abc123
# Step 2: review founder output, then approve
agentsuite pipeline approve --pipeline-id pipeline-20260501T120000-abc123
# Repeat approve for each subsequent agent, or check status at any time:
agentsuite pipeline status --pipeline-id pipeline-20260501T120000-abc123
# Lost the pipeline ID? List all pipelines:
agentsuite pipeline listAgents requiring extra inputs (engineering, trust_risk, cio) need a JSON file passed via --agent-inputs:
// agent-inputs.json
{
"engineering": {
"tech_stack": "Python + FastAPI",
"scale_requirements": "1k RPM, 99.9% uptime"
}
}agentsuite pipeline run \
--agents founder,engineering \
--project-slug my-startup \
--business-goal "Launch a B2B invoicing tool" \
--agent-inputs agent-inputs.json \
--auto-approveMCP (Claude Code / Cowork / Codex / Antigravity): use agentsuite_pipeline_run, agentsuite_pipeline_approve, agentsuite_pipeline_status.
| Env var | Default | Purpose |
|---|---|---|
AGENTSUITE_ENABLED_AGENTS |
founder |
Comma-separated agent names to expose (e.g. founder,design,product,engineering) |
AGENTSUITE_OUTPUT_DIR |
.agentsuite |
Where artifacts are written |
AGENTSUITE_LLM_PROVIDER |
(auto-detect) | Force anthropic, openai, gemini, or ollama |
AGENTSUITE_COST_CAP_USD |
5.0 |
Hard kill cap per run |
AGENTSUITE_EXPOSE_STAGES |
(off) | Set true to expose founder_intake/extract/spec/execute/qa as MCP tools |
|
A real agentsuite founder run end-to-end. Five stages, JSON summary. |
 |
The 14-artifact run directory after success. |
 |
brand-system.md rendered. |
 |
qa_report.md rendered. |
 |
_kernel/<project_slug>/ after approval. |
Browse a complete sample run on GitHub at examples/sample-output/founder/ — every artifact a Founder run produces, with the same on-disk shape as a live run; the free-text bodies are deterministic-mock scaffold (see the directory's README for what's authentic vs. scaffold).
AgentSuite runs every request through a five-stage pipeline:
flowchart LR
A[Intake] --> B[Extract]
B --> C[Spec]
C --> D[Execute]
D --> E[QA]
E -->|score ≥ 7.0| F[✅ Approved]
E -->|score < 7.0| G[🔄 Approval Gate]
G -->|human approves| F
Each stage is handled by the same kernel regardless of which agent is running. The QA stage scores output against a 9-dimension rubric; runs scoring below 7.0 enter an approval gate before artifacts are written.
┌───────────────┐
│ Harness │ (Codex / Claude Code / Cowork)
└─────┬─────────┘
│ stdio MCP
┌─────▼─────────┐
│ agentsuite-mcp│
└─────┬─────────┘
│
┌───────────┴───────────┐
│ │
┌─────▼──────┐ ┌──────▼──────┐
│ Kernel │◄────────│ FounderAgent│
│ │ │ │
│ schema/ │ │ stages/ │
│ qa/ │ │ rubric.py │
│ cost/ │ │ templates/ │
│ artifacts/ │ │ prompts/ │
│ approval/ │ └─────────────┘
│ base_agent │
└────────────┘
Full architecture diagram with all agents: see docs/README-FULL.pdf.
Shipped:
- v0.1.0 — Specification Kernel + Founder Agent
- v0.2.0 — Design Agent (brief generation, brand QA scoring)
- v0.3.0 — Product Agent (PM intent → UI spec → coding handoff)
- v0.4.0 — Engineering Agent (architecture decisions, system design, API specs, security review, deployment, runbook, tech-debt register, performance requirements)
- v0.5.0 — Marketing Agent (campaign brief, audience profile, messaging framework, content calendar, channel strategy, SEO keyword plan, competitive positioning, launch plan, measurement framework)
- v0.6.0 — Trust/Risk Agent (threat model, risk register, control framework, incident response plan, compliance matrix, vendor risk assessment, security policy, audit readiness report, residual risk acceptance)
- v0.7.0 — CIO Agent (IT strategy, technology roadmap, vendor portfolio, digital transformation plan, IT governance framework, enterprise architecture, budget allocation model, workforce development plan, IT risk appetite statement)
- v0.8.0 — path-traversal validation, MCP tool naming fixes, RetryingLLMProvider hardening
- v0.9.0–v0.9.1 — engineering hardening sprint (676 tests, 7 ADRs, resume idempotency contract)
- v1.0.0–v1.0.8 — Sprint 2 audit + remediation; 1066 tests green; cleanroom PASS
- v1.0.9 — Sprint 3 audit + remediation; 12 Nit findings resolved; full 4-sprint audit cycle complete
- v1.0.10 — Sprint 4 — 12 Nit findings resolved; full 57-finding audit cycle complete
- v1.0.11 — Sprint 5 — audience messaging refresh; seven-agent parity
- v1.0.12 — Sprint 6 — multi-agent pipeline orchestration (
pipeline run/approve/status); progress feedback - v1.0.13 — Sprint 7 — artifact visibility before approve;
pipeline list; landing page overhaul - v1.1.0 — K1 cross-stage context accumulator; K2 intra-stage progress events (
stage_progress,context_update) inBaseAgent.run()andPipelineOrchestrator
Roadmap (v1.2.x candidates):
- Per-day cost cap
- 8th agent TBD
- USER-MANUAL.md — plain-language walkthrough
- README-FULL.pdf — full reference with architecture diagrams
- CONTRIBUTING.md — dev setup + agent-implementation guide
- CHANGELOG.md — release notes
MIT — see LICENSE.