AnveInspect: the open-source core of ThoughtShieldAI

See every AI agent you run — what it costs, who spawned what, and get paged when one silently stops.

The open-source core powering ThoughtShieldAI — the AI control plane for enterprise Agentic AI. Everything in this repo is real, shipped, and MIT-licensed; nothing here is a mockup.

Teams now run dozens of coding agents across Claude Code, Codex, OpenClaw, Hermes, and their clouds — and nobody can answer "what's running, why, and what did it cost?" AnveInspect answers it with zero instrumentation: it reads the logs your agents already write and reuses the platform CLIs you're already signed into. One dashboard, one CLI, and an MCP server so Claude itself can operate the fleet.

git clone https://github.com/ANVEAI/anveinspect && cd anveinspect && npm install   # npm publish pending — then just: npx anveinspect
npx anveinspect doctor       # detects platforms, reuses your CLI logins, first scan
npx anveinspect dash         # the fleet, live at localhost:4177
npx anveinspect --help       # every command, one line each

30 seconds from install to a full inventory — no API keys pasted, no SDK added to your agents, no code changed.

platform CLIs (gcloud/aws/wrangler/az)  ─┐
claude/codex/openclaw/hermes local logs ─┼─▶ collector ─▶ SQLite ─┬─▶ CLI
claude hooks ─▶ fleet-emit ─▶ spool ─────┘                        ├─▶ MCP (Claude operates it)
                                                                  ├─▶ dashboard (:4177)
                                          cadence engine ─▶ alerts ┴─▶ Slack / macOS (the pager)

Connect your AI tools (bindings)

One command prints the exact registration for each tool, resolving your install's paths:

anveinspect setup claude    # Claude Code: plugin dir or `claude mcp add` one-liner
anveinspect setup codex     # Codex: ~/.codex/config.toml block
anveinspect setup cursor    # Cursor: .cursor/mcp.json block

doctor reuses the sessions you already have — gcloud auth login, aws configure, wrangler login, az login — so cloud platforms connect with no API keys to paste. Anything not signed in prints the exact one-line command to fix it. Example:

✓ Claude Code            reads ~/.claude/projects (no signin needed)
✓ Codex                  reads ~/.codex/sessions (no signin needed)
✓ Google Vertex          gcloud authed, project my-proj
→ Cloudflare Workers     wrangler present but not logged in
      sign in:  wrangler login

Then:

anveinspect status      # fleet pulse + open alerts + stale agents
anveinspect report      # full AI-ready briefing
anveinspect dash        # dashboard at http://localhost:4177

Operated by Claude (the primary interface)

Install the plugin (claude --plugin-dir packages/plugin) and any Claude session becomes your fleet operator. Ask it: "connect my platforms" → /anveinspect:setup; "how's my fleet?" → a full briefing; "watch my nightly agent, it runs at 3am" → it declares the cadence and the next silent failure pages you.

• 16 MCP tools — fleet_setup, fleet_status, fleet_attention, fleet_agents, fleet_agent_detail, fleet_scan, fleet_declare_cadence, fleet_check, fleet_ack, fleet_report, fleet_insights, fleet_lineage, fleet_analytics, fleet_tag, fleet_connectors_sync, fleet_connectors_status — zod-validated, structured output, read-only annotations.
• Skill fleet-ops + slash commands /anveinspect:setup, :status, :scan, :attention, :watch, :brief.
• Hooks capture live run events for crash detection and trigger classification.

Dashboard (enterprise control plane)

Six views (npm run dash → http://localhost:4177):

• Overview — KPIs (agents, alerts, stale, tokens, est. cost), token-usage trend, agents-by-platform, estimated-cost breakdown (by model + most-expensive agents), activity-by-hour, top consumers.
• Agents — full inventory: live search, vendor/status filters, sortable columns, inline tag chips; click any row for the detail drawer (runs, cadence, token sparkline, run-history timeline, tag editor).
• Activity — live chronological feed of runs across every platform.
• Lineage — the agent relationship graph (interactive force layout: who spawns whom across all executions, node size = run volume, edge weight + arrow = spawn direction and count, drag/hover/click-through). Double-click any agent to open it as a mindmap: the agents it spawns unfold to its right, the agents that spawn it to its left, a level at a time (⊕ badges say how many more); double-click again to fold a branch away. Plus the spawn-tree explorer for single executions (click a root to expand, with subtree tokens and descendant counts).
• Attention — open alerts (with Ack) + stale agents.
• Connectors — per-platform plug-and-play health.

Insights & cost

anveinspect insights (analytics JSON), anveinspect costs (estimated 30-day spend by model / agent, busiest hours, failure bursts). Cost estimates come from an editable ~/.anveinspect/pricing.json; token counts are exact, dollars are estimates, and every result reports priced/unpriced/tokenless coverage.

Control

anveinspect tag add|remove|list <agent> <tag> — organize the fleet with local tags/groups (cohort, tier, owner, "retire"). Purely local metadata: never touches a platform or changes agent behavior.

Silent-failure alerting (the wedge)

Declare when an agent should run and get paged when it doesn't:

This is the beginning of the README. Read the rest at the original. Read the full README ↗