Version:
3.3.0Status:ActiveUpdated:2026-02-16Module:root
Deterministic agents for auditable work.
Project page: https://lasteld.github.io/AMS/
AMS is an MCP server that adds memory, audit integrity, and deterministic orchestration to AI-assisted development workflows.
AMS is an Agent Management System for teams that want agent output to be:
- Stateful across sessions (not "chat-only memory")
- Auditable and tamper-evident
- Reproducible through explicit orchestration rules
- Safe to run with rollback and access controls
At runtime, AMS combines:
- Unified MCP tools (
projects/unified-mcp) - GSD pipeline integration (
projects/get-shit-done) - Educational roadmap system (
data/roadmaps) - Optional product layer UI/bridge (
projects/ams-mcp-ui)
Typical agent workflows fail in the same ways:
- No durable memory across sessions
- No reliable proof of what was changed and why
- Orchestration drift ("half works, half breaks")
- Unsafe file mutation without recovery path
- Weak onboarding because architecture intent is not explicit
AMS exists to solve those issues with one system:
- Memory frames + RAG indexing
- Action hash chain + Merkle audit proofs
- Deterministic phase pipeline
- Git checkpoint and rollback controls
- RBAC/ACL for project-scoped operations
- Install dependencies:
cd projects/unified-mcp
npm install- Configure MCP client:
{
"mcpServers": {
"ams": {
"command": "node",
"args": ["C:/AMS/projects/unified-mcp/src/server.js"],
"env": {
"AMS_ROOT": "C:/AMS"
}
}
}
}- Run smoke commands in your MCP client:
unified_vitals
unified_help
roadmap_list
If these three succeed, server routing, database, and roadmap loading are operational.
For release-grade validation in projects/unified-mcp:
npm test
npm run arch:gate
npm run pillars:gate:deepAMS uses two phase models.
- Execution pipeline for autonomous runs:
INIT -> GATHER -> ANALYZE -> PLAN -> [DRY_RUN|APPLY] -> VERIFY -> DONE
- Audit integrity phases:
- Phase 1: Context snapshots with SHA-256 verification
- Phase 2: Action hash chain
- Phase 3: Thought chain (decision trace)
- Phase 4: Merkle tree finalization
- Phase 5: Proof verification and audit export flow
Current status:
- Core phases are implemented and wired in the MCP server.
- Product/UI layer is available as a separate project (
projects/ams-mcp-ui) and evolves independently.
- Tool profiles and token economy:
AMS_TOOL_PROFILE=full|default|core|customAMS_TOOL_PROFILE_STRICT=true|false
- Orchestration dependency graph:
- Mandatory pre/post chains (A -> B, A -> C, A -> D) are encoded in graph + workflow layer.
- Memory and retrieval:
analysis_rag_index,analysis_rag_search,memory_pack,memory_verify
- Safety controls:
- Circuit breaker, retry/backoff, rate limiting, rollback
- Access control:
- Project-scoped ACL via
AMS_USER_IDandacl_*tools
- Project-scoped ACL via
AMS/
|- projects/
| |- unified-mcp/ # Main MCP server (controllers/domains/db/middleware)
| |- get-shit-done/ # GSD planning/execution framework
| `- ams-mcp-ui/ # Optional bridge server + dashboard UI
|- data/
| |- ams.db # SQLite runtime database
| |- roadmaps/ # Learning roadmap content
| `- state/ # Runtime mutable state files
|- docs/ # Stable docs, planning, archive, ideas
|- scripts/ops/ # Operational verification/maintenance scripts
`- config/ # Environment and integration configs
For developer onboarding and code-review orientation, start with docs/reference/DEVELOPER_GUIDE.md.
docs/reference/DEVELOPER_GUIDE.md- onboarding, architecture walkthrough, review flowSETUP_GUIDE.md- installation and MCP client setupUSER_GUIDE.md- practical day-to-day tool usageDEPLOYMENT.md- production deployment and release checksdocs/reference/PIPELINE_USAGE_GUIDE.md- autonomous pipeline operationdocs/reference/AMS_QR_MEMORY.md- AMS-QR memory modeldocs/reference/COMMANDS.md- tool catalogARCHITECTURAL_ANALYSIS.md- deep architectural trade-offs
- Version:
3.3.0 - Tools: use
unified_helpfor live surface count - Roadmaps: use
roadmap_listfor live count