</> mad-agentic.io Open Docs

AGENT WORKSPACE / CONTROL CENTER

Run autonomous workflows with production-grade control.

mad-agentic.io gives teams a single surface for run state, tool traces, memory context, approvals, and security boundaries.

Open Control Surface Read Documentation
Running agents
28
Tool calls / day
4.2M
Approval latency
96ms
Live Workspace cluster: prod-apac / fabric-03
Run success 99.4%
Guardrail hits 14
Escalations 3
agent-runs real-time execution monitor
Queued 12
Running 07
Blocked 02

System health

  • research-agentCompleted / 3 findings
  • release-orchestratorAwaiting approval
  • support-triageEscalated to operator

Connected to the stack your agents already use

GitHub GitLab Docker OpenAI Slack Jira

WORKSPACE SIGNALS

Core control primitives for high-trust autonomous systems.

Agent Runs

See queue depth, runtime state, retries, failures, and handoffs in one panel.

Tool Calls

Inspect external calls with outcomes, latency, and policy context.

Memory

Separate session, repository, and policy memory for cleaner, auditable reasoning.

Approval Flow

Route sensitive actions through human checkpoints before execution.

ACTIVITY TIMELINE

Track what happened, why it happened, and who approved it.

Chronology, checkpoints, and evidence are first-class UI elements.

Observed
Approval
Executed
run-2481 / operational timeline support-triage > billing-escalation
  1. 09:12
    Intake received

    Issue classified and prior support context loaded.

  2. 09:13
    Tool chain executed

    CRM lookup, billing fetch, and notifier completed within budget.

  3. 09:14
    Approval requested

    Refund action crossed policy boundary and paused for review.

  4. 09:16
    Action released

    Operator approved partial refund; workflow closed with audit entry.

MULTI-AGENT ORCHESTRATION

Compose specialist agents into one governed execution graph.

See orchestration notes
01Research Agent

Collects context, reads memory, builds the initial operating brief.

02Planner Agent

Converts objectives into bounded policy-aware action plans.

03Executor Agent

Runs tool chains and publishes verifiable state changes.

04Reviewer Agent

Scores risk and routes uncertain steps to human review.

05Memory Agent

Persists validated outcomes to session and repository memory.

06Audit Agent

Records actor, tools, approvals, and final decision trail.

USE CASES

Deploy the same control model across high-impact workflows.

Pick a lane, then reuse the same run, approval, and audit primitives.

Engineering

Release Assistant

Plan rollout, run checks, open approvals, and publish release artifacts with full traceability.

Support

Ticket Triage Agent

Classify issues, fetch customer context, draft responses, and escalate only risk-bound actions.

Operations

Incident Response Agent

Correlate signals, recommend runbooks, execute safe steps, and request approval on critical paths.

Growth

Experiment Operator

Generate variants, sync analytics tools, and maintain a clean audit trail for each release decision.

GITHUB PRODUCTS

Quick start guidance for mad-agentic GitHub products.

View all repositories

ProxyAPI.MAD

Open repository

Admin dashboard plus proxy backend for AI CLI tools, compatible with OpenAI/Gemini/Claude/Codex and optimized for local, self-managed provider setups.

Quick Start (Windows)

  1. Clone the `ProxyAPI.MAD` repository.
  2. From project root, run `run-dev.bat` to start backend and frontend.
  3. Or run `run-real.bat` for backend-only mode.
Backend: Go (`proxyapi_core/`) Frontend: React + Vite + TypeScript (`frontend/`)

SECURITY / AUDIT / GUARDRAILS

Keep autonomous actions bounded before they touch real systems.

Guardrails and audit table
Control How mad-agentic.io handles it Operator outcome
Policy guardrails Agents evaluate scope, cost, and sensitivity before execution. Unsafe actions are blocked early.
Approval routing Human gates appear inline when an action crosses policy thresholds. Operators approve only the narrow action that matters.
Audit logging Each tool call, memory write, and approval is recorded to a visible trail. Investigations are faster and evidence-ready.
Execution boundaries Runs are limited by environment, tool class, and action type. Teams start safe and scale autonomy gradually.

Audit Log

  • 09:14:06refund.action queued
  • 09:14:08policy.guardrail matched
  • 09:16:11human approval granted

Guardrail Pack

  • Tool budget limits by workflow type
  • Restricted actions on production resources
  • Mandatory approval for destructive or financial actions

DOCUMENTATION

Need implementation detail beyond the UI layer?

Open repository docs to maintain this control-center page and evolve the operating model.

Open README Documentation