Skip to main content
Documentation

TalkOps Overview

TalkOps is an autonomous DevOps operations framework built on LangGraph and the TypeSafe AI Jev engine. It executes end-to-end operational workflows across multi-cloud infrastructure (AWS, Azure, GCP), Kubernetes clusters, Infrastructure as Code (Terraform, OpenTofu, CloudFormation, Ansible), CI/CD pipelines, observability stacks, FinOps, and application repositories.

Deploy TalkOps locally in your terminal (talkops), headlessly inside CI/CD pipelines (talkops -p), or centrally as a multi-tenant server (talkops serve) that listens on production webhooks (/webhooks/{provider}) and serves multiple connected operator clients (--server-url).

⚡ Jev Smart Mode (Cost, Latency & Safety)

Routes prompts across 3 model tiers and reasoning levels in <70ms (slashing token cost by 80–90% on reflex lookups with effort="off") and classifies tool blast radius in <100ms with zero LLM prompt token overhead.

🌐 Central Server & Multi-Client Runtime

Run talkops serve with PostgreSQL state persistence, 1-step API-key-to-JWT token exchange, and multi-user thread isolation so multiple engineers can connect via --server-url.

🔔 Event-Driven Webhooks & Write-Backs

Listen on 9 built-in webhook adapters (PagerDuty, Alertmanager, Datadog, AWS SNS, CloudEvents, ServiceNow, Jenkins, GitHub, Azure Monitor) and post automated RCAs back to tickets, PRs, and Slack.

🤖 Dynamic DevOps Subagents

Delegate complex investigations in parallel across dynamic plugin subagents, project (.talkops/agents/), user, and async remote subagents with independent Smart Mode routing.

What TalkOps can do across DevOps operations​

TalkOps is not limited to writing code or running single-cloud CLI commands. Anything practically achievable by a senior SRE, platform engineer, or cloud architect from a terminal or API can be orchestrated, verified, and governed through TalkOps:

  • Autonomous Incident Triage & RCA: Receive real-time webhooks from PagerDuty, Prometheus Alertmanager, Datadog, AWS SNS, or ServiceNow; deduplicate alert storms; correlate metrics, logs, and recent deployments in a stateful thread; and write the Root Cause Analysis back to the incident ticket or Slack.
  • Multi-Cloud Operations (AWS, Azure, GCP): Inspect and remediate EC2/EKS/RDS/IAM on AWS, AKS/VNet/Monitor on Azure, and GKE/Cloud Run/IAM on GCP using native SDKs and CLI credential chains.
  • Kubernetes & Container Orchestration: Diagnose CrashLoopBackOff, OOMKilled, PVC binding failures, CoreDNS latency, and Ingress/NetworkPolicy misconfigurations; audit and upgrade Helm releases.
  • Infrastructure as Code & GitOps: Author, validate, plan, and refactor Terraform, OpenTofu, Terragrunt, AWS CloudFormation, and Ansible playbooks; reconcile ArgoCD/Flux drift.
  • CI/CD Pipeline Recovery & Code Review: Analyze failed GitHub Actions workflows and Jenkins builds, fix underlying code or pipeline YAML issues, run diff-aware /review checks, and comment directly on pull requests.
  • FinOps Cost Optimization & Security Auditing: Identify unattached volumes, idle load balancers, and over-provisioned compute; audit IAM policies for least-privilege compliance.

Core architecture​

1. Three deployment topologies​

TalkOps runs the exact same LangGraph agent architecture across three operational modes:

TopologyHow to LaunchBest For
Local Interactive TUItalkops --smartIndividual engineers diagnosing clusters, writing IaC, or reviewing code locally with SQLite persistence (~/.talkops/sessions.db).
Headless CLI & CI/CDtalkops -p "<prompt>" --smart --criteria "..."Automated CI/CD gates, drift detection cron jobs, and scripted remediation loops with --max-turns and rubric evaluation.
Central Server + Multi-ClientServer: talkops serve --postgres-uri ...
Client: talkops --server-url http://host:2024		Platform teams sharing threads, ingesting 24/7 production webhooks, and enforcing centralized JWT auth and PostgreSQL persistence.

2. Jev Smart Mode: Dual-engine cost, latency & safety optimization​

When running in Smart Mode (--smart, /smart, or Shift+Tab), TalkOps uses the TypeSafe AI Jev engine in two distinct places on every turn:

  1. Dynamic Model & Reasoning Router (<70ms via JevDynamicModelRouterMiddleware):
    • Classifies the incoming prompt and routing_state (agent_role, active_capabilities, available_subagents, workspace_profile) in <70ms.
    • Routes Tier 0 (fast) queries (greetings, status checks, single-item lookups) to lightweight models (haiku, mini, flash-lite) with effort = "off"—eliminating reasoning latency and reducing token cost by 80–90%.
    • Routes Tier 1 (standard) single-domain DevOps and coding tasks to balanced models (sonnet, gpt-5.4, flash) with effort = "medium" or "low".
    • Routes Tier 2 (powerful) complex tasks (multi-agent delegation, SRE incident root-cause correlation, FinOps optimization, security audits, cross-service refactoring) to frontier models (opus, o3-pro, gemini-3.1-pro) with effort = "max" or "xhigh".
    • Includes confidence escalation safeguards (escalating to powerful if confidence < 0.55 on high-complexity markers) and automatic baseline restoration when switching back to /manual or /auto.
  2. System One Tool Safety Classifier (<100ms via JevSecurityClassifier):
    • Evaluates every proposed shell command, file edit, MCP tool call, and subagent dispatch in <100ms with zero LLM prompt token overhead across is_mutating (Noul), blast_radius (Score 0–2), and intent_alignment (Noul).
    • Auto-approves aligned read-only and controlled operations while failing closed to human approval on production mutations or critical blast radius (>= 1.2).

3. Dynamic subagent orchestration​

Instead of overloading a single context window, the TalkOps orchestrator delegates domain-specific work to dynamic subagents (src/talkops/subagents/loader.py) with isolated memory (SubagentMemoryStore), scoped skills, and dedicated SUBAGENT_ROUTE_CRITERIA Smart Mode routing:

Subagent SourcePath / ConfigPurpose
Agent Plugins~/.talkops/plugins/<plugin>/agents/*.mdDomain specialist packs installed via talkops plugin install (e.g., AWS SRE, FinOps, EKS, Security, Observability).
Project Subagents.talkops/agents/*.md or .agents/*.mdVersion-controlled repository and infrastructure specialists shared across the team.
User Subagents~/.talkops/agents/*.mdPersonal workstation subagents available across all repositories.
Async Remote Subagents[async_subagents] in config.tomlBackground LangGraph subagent graphs running on remote endpoints.

Documentation map​

Getting Started​

  • Quickstart — Install TalkOps, configure credentials, and run your first local, headless, or remote server session
  • CLI & Slash Command Reference — Dedicated mini-sections for every CLI subcommand (serve, auth, threads, mcp, plugin, skills, doctor) and interactive /slash command

Core Operations, Server & Safety​

Configuration & Providers​

Customization & Extensibility​

  • Subagents — Dynamic plugin subagents, custom .talkops/agents/*.md definitions, and Smart Mode subagent routing
  • Memory & Skills — Persistent AGENTS.md memory, reusable SKILL.md operational runbooks, and cryptographically verified scripts
  • Custom Commands — Creating project and user slash commands in .talkops/commands/
  • Plugins — Installing and packaging multi-component DevOps plugins
  • MCP Tools — Attaching external stdio, sse, and http Model Context Protocol tool servers
  • Hooks — Running deterministic scripts on SessionStart, PreToolUse, PostToolUse, and webhook.* lifecycle events
  • Server Add-ons — Extending talkops serve with modular FastAPI routers, background workers, and event bus subscribers
  • Server Add-on Developer Guide — Protocol reference and step-by-step guide for authoring custom server add-ons

Agents, Plugins Marketplace & Open-Source MCP Servers​

  • OpsCode Terminal Coding Agent — Open-source terminal AI coding agent (ops) for Terraform, OpenTofu, Kubernetes, and cloud repositories
  • Autonomous Kubernetes Agent — Extensible Kubernetes operations agent (k8s-autopilot) for Helm, GitOps, and observability
  • TalkOps DevOps Plugins Marketplace — 40 production plugins: 17 AWS specialist sub-agents, 21 vertical capability bundles (245 skills, 109 commands), and 3-layer IAM/mutation safety guardrails
  • Open-Source DevOps MCP Servers — 10 dedicated Model Context Protocol servers for Helm, ArgoCD, Argo Rollouts, Prometheus, Alertmanager, OpenTelemetry, Tempo, Loki, Traefik, and Terraform
On this page