Proofpane — Architecture Map

A multi-tenant AI governance + multi-agent orchestration platform with compliance-by-construction: every AI call passes a policy gate, lands in a hash-chained audit log, and exports as a signed Evidence Pack. Ships in four deployment shapes (cloud-only · cloud+tray · cloud+daemon · local-MCP) — see below.

Deployment shapes

One codebase, four deployment topologies. Each mode is real today — the daemon binary, the tray app, and the cloud are all one repo.

Mode 1

Cloud-only (SaaS)

Web user, no install

Browser→Cloud (FastAPI + SQLite + Qdrant)

All execution server-side. Tool runtime = LocalToolRuntime in the cloud process. Fastest path to demo.

Mode 2

Cloud + Tauri Tray

Web user + native HITL approvals

Browser→Cloud→Tauri Tray (SSE+REST, OS keyring JWT)

Tray pops native HITL approval dialogs when cp.tool.hitl fires; auto-approve gate uses precedent engine. Cross-instance reconcile via cp.hitl.resolve.

Mode 3

Cloud + Paired Daemon

Power user with local files / tools

Browser→Cloud→Daemon `run` (WebSocket reverse RPC)

Cloud → user machine: 7 local tools (read/write/edit/bash/glob/grep/list) routed via RemoteToolRuntime. Pairing code → device token bound to caller's org.

Mode 4

Local-first MCP (no cloud)

Claude Desktop / Cursor / Continue

MCP host→Daemon `mcp` (stdio MCP server)

Same daemon binary, mcp role. Tools exposed as MCP. No governance overlay in this mode — pure tool execution. Set CP_VENDOR=claude_desktop / codex_desktop for vendor tagging.

Request lifecycle

One LLM call traverses all 7 gates. Failure at any gate is itself audited.

Request

User / workflow / MCP client

›

Auth + Tenant

JWT → org_id binding

›

Policy gate

runtime_policy + DLP

›

Budget gate

per-org USD cap

›

LLM call

5 providers + composed Capsule

›

Audit + Cost

hash-chained · token rollup

›

Evidence Pack

Ed25519-signed export

Manage — governance objects

AI Use Cases

Per-org case intake + RAG

· AI-governance-shaped intake form (post-cleanup)
· Filesystem store: data/projects/{org_id}/{case_id}/
· Qdrant per-case collections · hybrid (BM25 + vector) search
· Workflow runs link back to a case for traceability

AI Use Cases →

Workflows & Skills

24 skills · 8 categories · visual builder

· Intake · risk · vendor · policy · monitoring · audit · training · external
· React Flow canvas · drag-drop palette
· AI Builder chat: proposes / edits the canvas
· Node types: skill · trigger · multi-agent · mcp_call
· Per-run HITL checkpoints with approval drawer
· is_template=1 = global catalog; everything else per-org

Workflows →

MCP Setup

Plug Claude/Cursor/etc into governance

· Daemon pair: short code → device token bound to org
· Org-scoped MCP server registry · stdio + http transports
· Publish lifecycle: draft → published → deprecated
· Members only see published; admin curates

MCP Setup →

MCP Activity

At-a-glance daemon trace

· Vendor-tagged ingestion (Claude Desktop · Codex Desktop)
· Cross-boundary block / HITL / allow rates · DLP masking events
· Per-agent policy scopes — code_agent vs hr_agent isolation
· Servers callable from workflows + skills

MCP Activity →

Watch — visibility dashboards

Mission Control (Audit)

Tamper-evident decision log

· Hash-chained audit_log (SHA-256, org_id in hash)
· 50+ event types: skill · workflow · MCP · multi-agent · tray · cp.*
· Mission Control: agents · terminal pool · HITL queue · precedent engine
· Live SSE event stream + CSV export bracketed by filters
· Evidence Pack: Ed25519-signed zip + offline verifier

Mission Control →

Cost & Budget

Spend gate + anomaly + forecast

· Per-org monthly USD cap · refusal is itself audited
· Anomaly flag: call > Nx recent average
· Month-end forecast (linear / rolling7d / worst_case + 10%)
· Live LiteLLM pricing fallback for unknown models (D33)
· Dashboard: KPIs · 30-day sparkline · top spenders

Cost dashboard →

Quality Stack

Hallucination + drift signals

· L3 closed-set membership against 259-control truth set
· L4 LLM-as-judge with grounding dimension
· L5 drift alerts: pass-rate + hallucination rate, suppressed independently
· Workflow eval suites (DAG node-output / trace-path / step-budget)
· Quality sampling daemon · hallucination triage queue

Quality dashboard →

Agent Runs (unified)

Tree-view across 4 agent kinds

· Workflow / Remediation / Multi-agent / Cross-vendor — one tree-view
· parent_run_id nests sub-agents (doc 16 contract)
· Live tree via SSE · pop-out window for side-by-side
· Index: 24h / 7d / 30d filter · kind chip · descendant rollup
· audit_log IS the state machine — view is pure projection

Agent runs →

Cross-vendor Eval

Same case, N providers, judged

· Pairwise diff on decision fields + schema
· Severity grades {info, warning, critical}
· Optional independent judge for semantic equivalence on non-critical pairs
· Live 4-level tree: suite → case → provider → judge (doc 16)

Cross-vendor →

Lab — self-evolution experiments

Self-Evolution (Dreams)

Gene · Capsule · approval loop

· Gene library (proposed / approved / deprecated · 400-char cap)
· Capsules: frozen composed prompts · exactly-one-active per (org, skill)
· Dream loop: audit signals → Gene proposals (HITL gate, never auto-approve)
· Sandbox replay (K samples) before approve
· Auto-deprecate: approved Genes never picked > stale_after_days
· META loop: code-agent's own system prompt is itself a Capsule

Dreams →

Parallel Universe

Exogenous dreamer · 5 sources

· 5 sources: anthropic-sdk-python · openai-python · arxiv cs.AI · GitHub Search · Hacker News
· Sonnet triager scores relevance · proposes Gene or code-agent
· Per-org budget meter · scheduled loop · stale auto-sweep
· Promote → HITL-gated downstream (Gene proposal / code-agent session)

Parallel Universe →

Repo Coder — Mode A

Dual role: tool + self-improvement executor

· Role 1 — general LLM coder: any goal, edits existing repo via git worktree
· Role 2 — self-improvement executor: Parallel Universe promotes here for code changes
· Disposable branch · active-cap 10 · 60s command timeout
· LLM tool loop: read / write / run / done
· Iter + token budgets (default 20 iter / 200k tokens)
· HITL gate: promote creates session, operator manually clicks Run loop
· META loop: own system prompt is itself a Gene-evolved Capsule
· Internal module + route still named code_agent — only the UI label changed

Repo Coder →

Greenfield — Mode B

Docker sandbox · new project

· Docker container · --network=none · isolated artifact build
· On Approve: tarball download · on Discard: container teardown
· Same AgentRunView tree-view as Repo Coder
· Mode A vs Mode B chosen by goal shape (edit existing vs greenfield)

Greenfield →

Admin — identity & access

Identity & Tenancy

Who's allowed to do what

· JWT auth · register / login
· Magic-link invitations (SMTP-delivered)
· Roles: owner / admin / member · destructive ops admin+
· Org switcher in account drawer
· Every resource scoped by org_id (SQL + filesystem)
· Cross-org tampering detected via hash chain

Users →

Foundation — stack & integrations

Compliance Frameworks

59 controls (truth-set: 259 IDs)

· NIST AI RMF · ISO 42001 · EU AI Act · GDPR · SOC 2 TSC
· Skill ↔ control mapping out of the box
· Per-org overrides (add / remove / approve)
· Closed-set ID truth-set drives hallucination guard (doc 15)

Compliance →

Multi-Agent Primitives

Consensus + adversarial review

· consensus(N agents, method=vote | entropy)
· adversarial_review: drafter → critic → reviser rounds
· REST endpoints AND workflow node types
· Semantic-entropy escalation for HITL routing
· Sub-agents emit own audit tree via parent_run_id

Cloud Runtime

FastAPI · SQLite WAL · Qdrant

· Single-process FastAPI (fly.io target)
· SQLite with WAL · tenant-scoped tables · Qdrant for RAG / precedents
· asyncio loops: scout · dream · workflow worker · drift sweep · staleness sweep
· Realtime: WS /devices/ws · SSE /audit/stream
· Alembic migrations (D29) · structured JSON logs (D30) · Sentry (D31)

Tauri Tray

Desktop HITL approval app

· Tauri 2 + React + Rust core · macOS / Windows / Linux
· Pair via short admin code → 90-day JWT
· OS keyring JWT (Keychain / Cred Mgr / Secret Service)
· SSE + REST: REST = source of truth, SSE = wakeup signal
· Auto-approve via precedent threshold; 'Approve & remember' (boost=3.0)
· Activity panel · cross-instance reconcile via cp.hitl.resolve

Hybrid Daemon

User-machine binary · two roles

· Pairing: short code → device token bound to org
· Role A — `run`: cloud-paired (WebSocket reverse RPC)
· Role B — `mcp`: stdio MCP server for Claude Desktop / Cursor / Continue
· 7 local tools: read / write / edit / bash / glob / grep / list
· Built as a single PyInstaller binary

LLM Providers

5 backends · one adapter

· Anthropic · OpenAI · DeepSeek · Google Gemini · OpenAI-compatible (vLLM / Together / Fireworks / Ollama)
· Per-org API keys + platform key fallback
· Tool-calling adapter normalised across vendors
· Schema sanitization for Gemini compatibility
· Vendor /models cache + LiteLLM live pricing fallback (D33)

External Integrations

Salesforce · Gmail · Slack

· Salesforce: REST + OAuth (password/refresh-token); governed lead intake skill
· Gmail OAuth + auto email processor
· Slack drift-alert notifier
· external.* skill prefix; PII safe — audits field NAMES not values

Documentation

Per-subsystem deep dives

· CLAUDE.md §4 — subsystem map + recipes + gotchas
· docs/architecture/01-…16 — one file per subsystem
· docs/architecture/15-quality-and-hallucination — 4-signal reference
· docs/architecture/16-agent-tree-view — parent_run_id contract

Stack

Frontend

· Next.js 14 (app router)
· Tailwind
· React Flow (builder)
· SSE streams

Backend

· FastAPI
· SQLite + WAL
· Qdrant (RAG / precedents)
· asyncio loops

Desktop

· Tauri 2 (Rust + WebView)
· PyInstaller daemon binary
· tauri-plugin-store
· OS keyring crate

Storage

· SQLite app.db (tenant-scoped)
· Filesystem cases
· Ed25519 signing key
· Audit hash chain

External

· Anthropic / OpenAI / DeepSeek / Gemini / OAI-compat
· GitHub · arxiv · HN
· Salesforce · Gmail · Slack
· MCP servers (stdio / http)

Pillar icons mirror the sidebar so this page reads like a map of the left-rail. When you ship a new capability, add it to a GROUP_* array in this file. See CLAUDE.md §4 and docs/architecture/ for subsystem deep dives.