MCP Fabric

A control plane for AI agents — register MCP servers, normalize tools into capabilities, control access with policies and trust levels, require human approval for sensitive actions, and audit everything.

MCP solves one important problem: a standard interface for AI tools and data. It does not solve the next problem that appears immediately after success: tool ecosystem sprawl.

MCP Fabric is that missing layer — the dashboard where platform teams connect, organize, control, review, and monitor every tool their agents use.

The Problem

Once a team adds multiple MCP servers, several hard questions emerge:

• Which tools should be exposed to which agents?
• How should overlapping capabilities be described?
• What trust level should be assigned to each server?
• Which agents get access to which tools — and who needs to approve?
• How should a platform team audit and govern usage across the whole tool ecosystem?

What's New in v0.3.0 — Schema-Digest Security

Full changelog: docs/CHANGELOG.md

v0.3.0 brings Schema-Digest Mappings — a security layer that detects when server tool schemas drift from their capability mappings. When a re-inspected server shows a changed tool schema, affected mappings are automatically marked stale and excluded from routing until an admin reviews and approves the change.

Key features:

• Schema-digest computation — SHA-256 hash of (tool_name + input_schema + output_schema) stored on each mapping at creation time
• Drift detection on re-inspect — schema changes automatically mark affected mappings as stale
• Digest-bound routing — only active mappings with matching digests are considered for routing
• OPA policy gates — deny_stale_mapping and untrusted_write rules in Rego policies
• Admin review UI — Pending Reviews page with approve/reject workflow and audit trail
• Trust Posture integration — "Pending Reviews" button linking to the review page

Key resources:

• Work Breakdown — Phase 14

Quick Start

# Install from PyPI
pip install mcp-fabric-toolmesh

# Clone and start the full stack (Docker)
git clone https://github.com/deghosal-2026/mcp-fabric.git
cd mcp-fabric
docker-compose up -d

# Or run locally without Docker:
poetry install && cd ui && npm install && cd ..
poetry run uvicorn api.main:app --reload &
cd ui && npm run dev

Documentation

What MCP Fabric Does

MCP Fabric is a control plane for AI agents. It sits between your MCP servers and the agents that consume them, giving platform teams a single dashboard to manage the full lifecycle:

1. Register servers → 2. Define capabilities → 3. Set policies
→ 4. Create agent classes + tokens → 5. Bundle into packs
→ 6. Assign packs → 7. Approve gated actions → 8. Audit everything

Architecture

Agent → MCP Fabric API → Registry → OPA Policy → Approval Gate → Target MCP Server
                            ↓
                      Audit Pipeline
                            ↓
                     ┌──────────────┐
                     │  Admin UI    │
                     │  (Dashboard, │
                     │   Config,    │
                     │   Monitor)   │
                     └──────────────┘

The Admin UI is the control plane. Platform teams use it to register servers, define capabilities, set policies, manage agent classes, bundle packs, review approvals, monitor alerts, and audit activity — all from a single dashboard.

Stack

• API: FastAPI
• Metadata store: PostgreSQL
• Cache: Redis
• Policy engine: OPA (Open Policy Agent) Rego policies
• Telemetry: OpenTelemetry, Tempo, Prometheus, Grafana
• UI: React
• Local dev: Docker Compose

Everything runs locally. No enterprise dependencies required.

Roadmap

v0.3.0 — Current:

• Schema-digest mappings: detect tool drift, mark stale, block routing until re-approved
• OPA deny rules: deny_stale_mapping, untrusted_write, raw_context for audit
• Admin review UI: approve/reject stale mappings with audit trail
• 326 automated tests (backend, UI, OPA, E2E) + 31 OPA policy tests

v0.4.0 — Planned:

• Multi-tenant scopes and namespace isolation
• Analytics and usage heatmaps
• Webhook integrations for external tooling
• Performance benchmarks and caching improvements

GA (v1.0.0) — Planned:

• Stabilization, security audit, production hardening, and enterprise features

Test Status

| Suite | Tests | Status | |---|---|---|---|---| | Backend unit (services, middleware, errors, models) | 326 | ✅ Passing | | OPA policy (Rego) | 31 | ✅ Passing | | UI unit/integration (Vitest) | 128 | ✅ Passing | | UI E2E + screenshots (Playwright) | 75 | ✅ Passing | | Docker Compose E2E (curl) | 6 | ✅ Scripts ready | | Total | 566 | |

make test        # Backend unit tests
make test-unit   # Unit tests only
cd ui && npm test  # UI tests
make opa-test    # OPA policy tests

Who It's For

• Platform teams managing multiple MCP servers
• Advanced agent builders who need governed tool access
• Developer experience teams building internal AI tooling platforms
• OSS builders creating reusable MCP ecosystems

License

MIT

See ROADMAP, GOVERNANCE, and CONTRIBUTING for project direction, governance, and how to contribute.