GitHub - klawsh/klaw.sh: kubectl for AI Agents

kubectl for AI Agents

Enterprise AI agent orchestration. Manage, monitor, and scale your AI workforce.
One binary. Deploys in seconds. Scales to hundreds of agents.

Documentation • What is klaw? • Quick Start • Slack Control • Architecture

What is klaw?

klaw is enterprise AI agent orchestration — like kubectl, but for AI agents.

# See all your agents
$ klaw get agents

NAME              NAMESPACE    STATUS    MODEL              LAST RUN
lead-scorer       sales        running   claude-sonnet-4    2m ago
competitor-watch  research     idle      gpt-4o             1h ago
ticket-handler    support      running   claude-sonnet-4    30s ago
report-gen        analytics    idle      claude-sonnet-4    6h ago

# Check what an agent is doing
$ klaw describe agent lead-scorer

Name:         lead-scorer
Namespace:    sales
Status:       Running
Model:        claude-sonnet-4-20250514
Skills:       crm, web-search
Tools:        hubspot, clearbit, web_fetch
Last Run:     2 minutes ago
Next Run:     in 58 minutes (cron: 0 * * * *)

# View real-time logs
$ klaw logs lead-scorer --follow

[14:32:01] Fetching new leads from HubSpot...
[14:32:03] Found 12 new leads
[14:32:05] Analyzing lead: john@acme.com
[14:32:08] Score: 85/100 (Enterprise, good fit)
[14:32:09] Updated HubSpot lead score
...

Or control everything from Slack

You: @klaw status

klaw: 📊 Agent Status
      ├── lead-scorer (sales) — running, 2m ago
      ├── competitor-watch (research) — idle
      ├── ticket-handler (support) — running, 30s ago
      └── report-gen (analytics) — idle

You: @klaw run competitor-watch

klaw: 🚀 Starting competitor-watch...
      Checking competitor.com/pricing...
      Found 2 pricing changes since yesterday.
      Posted summary to #competitive-intel

The Problem

You're running AI agents in production:

Lead Scorer — analyzes CRM leads every hour
Competitor Watch — monitors competitor websites daily
Ticket Handler — auto-responds to support tickets
Report Generator — creates weekly analytics reports

But managing them is chaos:

Challenge	Current State	With klaw
Visibility	"Is the agent running? What's it doing?"	`klaw get agents`, `klaw logs`
Isolation	Sales agent accessing support secrets	Namespaces with scoped permissions
Scheduling	Messy cron jobs, Lambda functions	`klaw cron create` — built-in
Scaling	Manual server provisioning	`klaw node join` — auto-dispatch
Debugging	grep through CloudWatch	`klaw logs agent --follow`
Deployment	Complex setup, many dependencies	Single binary, one command

OpenClaw works, but deployment is painful and scaling is worse.

klaw brings Kubernetes-style operations to AI agents. One binary. Deploys in seconds.

Quick Start

1. Install

curl -fsSL https://klaw.sh/install.sh | sh

2. Configure

# Pick one provider
export ANTHROPIC_API_KEY=sk-ant-...      # Direct Anthropic
export OPENROUTER_API_KEY=sk-or-...      # OpenRouter (100+ models)
export EACHLABS_API_KEY=...              # each::labs (300+ models)

3. Run

# Interactive chat
klaw chat

# Or start the full platform (Slack + scheduler + agents)
export SLACK_BOT_TOKEN=xoxb-...
export SLACK_APP_TOKEN=xapp-...
klaw start

Real-World Examples

Sales: Lead Scoring

# Create the agent
klaw create agent lead-scorer \
  --namespace sales \
  --model claude-sonnet-4-20250514 \
  --skills crm,web-search

# Schedule hourly runs
klaw cron create score-leads \
  --schedule "0 * * * *" \
  --agent lead-scorer \
  --task "Analyze new leads in HubSpot, score 1-100 based on fit, update Lead Score field"

# Check status anytime
klaw get agents -n sales
klaw logs lead-scorer

Research: Competitor Intelligence

klaw create agent competitor-watch \
  --namespace research \
  --model gpt-4o \
  --skills web-search,web-fetch,slack

klaw cron create competitor-daily \
  --schedule "0 9 * * *" \
  --agent competitor-watch \
  --task "Check competitor.com/pricing for changes. Post diff to #competitive-intel"

Support: Ticket Automation

klaw create agent ticket-handler \
  --namespace support \
  --model claude-sonnet-4-20250514 \
  --skills zendesk,slack

# This agent responds to Slack mentions automatically
# @klaw check ticket #12345
# @klaw draft response for angry customer

Analytics: Automated Reports

klaw create agent report-gen \
  --namespace analytics \
  --model claude-sonnet-4-20250514 \
  --skills sql,slack,charts

klaw cron create weekly-report \
  --schedule "0 8 * * MON" \
  --agent report-gen \
  --task "Query last week's metrics, generate summary with charts, post to #team-updates"

Slack Control

klaw turns Slack into your AI command center:

# Agent management
@klaw status                          # List all agents
@klaw describe lead-scorer            # Agent details
@klaw logs ticket-handler             # Recent logs

# Run tasks
@klaw run lead-scorer                 # Trigger immediately
@klaw ask lead-scorer "score this: john@bigco.com"

# Scheduling
@klaw cron list                       # View scheduled jobs
@klaw cron disable daily-report       # Pause a job

# Quick queries
@klaw "summarize today's support tickets"
@klaw "what did competitor-watch find yesterday?"

Architecture

┌────────────────────────────────────────────────────────────────────────┐
│                           KLAW CONTROL PLANE                            │
│                                                                          │
│  ┌────────────────────────────────────────────────────────────────┐    │
│  │                      NAMESPACES (Isolation)                     │    │
│  │                                                                  │    │
│  │  ┌───────────────┐  ┌───────────────┐  ┌───────────────┐       │    │
│  │  │    SALES      │  │   RESEARCH    │  │    SUPPORT    │       │    │
│  │  │               │  │               │  │               │       │    │
│  │  │ lead-scorer   │  │ competitor-   │  │ ticket-       │       │    │
│  │  │               │  │ watch         │  │ handler       │       │    │
│  │  │ ┌───────────┐ │  │ ┌───────────┐ │  │ ┌───────────┐ │       │    │
│  │  │ │ tools:    │ │  │ │ tools:    │ │  │ │ tools:    │ │       │    │
│  │  │ │ • hubspot │ │  │ │ • web     │ │  │ │ • zendesk │ │       │    │
│  │  │ │ • clearbit│ │  │ │ • slack   │ │  │ │ • slack   │ │       │    │
│  │  │ └───────────┘ │  │ └───────────┘ │  │ └───────────┘ │       │    │
│  │  │               │  │               │  │               │       │    │
│  │  │ secrets:      │  │ secrets:      │  │ secrets:      │       │    │
│  │  │ HUBSPOT_KEY   │  │ SLACK_TOKEN   │  │ ZENDESK_KEY   │       │    │
│  │  └───────────────┘  └───────────────┘  └───────────────┘       │    │
│  └────────────────────────────────────────────────────────────────┘    │
│                                                                          │
│  ┌─────────────┐  ┌─────────────┐  ┌─────────────┐  ┌─────────────┐    │
│  │  SCHEDULER  │  │  CHANNELS   │  │   ROUTER    │  │    NODES    │    │
│  │  (cron)     │  │(Slack, CLI) │  │ (300+ LLMs) │  │  (workers)  │    │
│  └─────────────┘  └─────────────┘  └─────────────┘  └─────────────┘    │
└────────────────────────────────────────────────────────────────────────┘

The Kubernetes Parallel

You Want To...	kubectl	klaw
List workloads	`kubectl get pods`	`klaw get agents`
Inspect	`kubectl describe pod`	`klaw describe agent`
View logs	`kubectl logs`	`klaw logs`
Deploy	`kubectl apply -f`	`klaw apply -f`
Isolate	Namespaces	Namespaces
Schedule	CronJob	`klaw cron`
Scale out	Add nodes	`klaw node join`

Deployment Modes

Single Node — Development & small teams

klaw chat              # Interactive
klaw start             # Full platform

Distributed — Production & enterprise

# Controller (central brain)
klaw controller start --port 9090

# Worker nodes join the cluster
klaw node join controller.internal:9090 --token $TOKEN

# Tasks auto-dispatch to available nodes
klaw dispatch "analyze all Q4 leads" --agent lead-scorer

CLI Reference

# Core
klaw chat                        # Interactive terminal chat
klaw start                       # Start platform (Slack + scheduler)
klaw dispatch "task"             # One-off task execution

# Agent Management (kubectl-style)
klaw get agents                  # List all agents
klaw get agents -n sales         # List in namespace
klaw describe agent <name>       # Detailed info
klaw create agent <name>         # Create new agent
klaw delete agent <name>         # Remove agent
klaw logs <agent>                # View logs
klaw logs <agent> -f             # Follow logs

# Namespaces
klaw get namespaces              # List namespaces
klaw create namespace <name>     # Create namespace
klaw config use-context <ns>     # Switch namespace

# Scheduling
klaw cron create <name>          # Create scheduled job
klaw cron list                   # List all jobs
klaw cron describe <name>        # Job details
klaw cron enable <name>          # Enable job
klaw cron disable <name>         # Disable job
klaw cron delete <name>          # Remove job

# Cluster (Distributed Mode)
klaw controller start            # Start controller
klaw node join <addr>            # Join cluster
klaw get nodes                   # List nodes
klaw drain node <name>           # Drain node

# Configuration
klaw config view                 # Show config
klaw config set <key> <value>    # Update config

Configuration

Minimal (`~/.klaw/config.toml`)

[provider.anthropic]
api_key = "${ANTHROPIC_API_KEY}"

Full Example

[defaults]
model = "claude-sonnet-4-20250514"
namespace = "default"

[provider.anthropic]
api_key = "${ANTHROPIC_API_KEY}"

[provider.eachlabs]
api_key = "${EACHLABS_API_KEY}"

[channel.slack]
enabled = true
bot_token = "${SLACK_BOT_TOKEN}"
app_token = "${SLACK_APP_TOKEN}"

[channel.cli]
enabled = true

[namespace.sales]
secrets = ["HUBSPOT_KEY", "CLEARBIT_KEY"]
allowed_tools = ["hubspot", "clearbit", "web_search"]

[namespace.support]
secrets = ["ZENDESK_KEY"]
allowed_tools = ["zendesk", "slack"]

[scheduler]
enabled = true
timezone = "America/New_York"

Custom Provider (OpenAI-Compatible)

Any OpenAI-compatible API can be used as a provider by setting base_url and model in config. This works with Ollama, LM Studio, vLLM, GLM, Minimax, Together AI, and more.

# Ollama (local)
[provider.ollama]
base_url = "http://localhost:11434/v1"
model = "llama3.2"

# LM Studio (local)
[provider.lmstudio]
base_url = "http://localhost:1234/v1"
model = "qwen2.5-coder-32b"

# Together AI
[provider.together]
api_key = "your-api-key"
base_url = "https://api.together.xyz/v1"
model = "meta-llama/Llama-3.3-70B-Instruct-Turbo"

# Any OpenAI-compatible endpoint
[provider.custom]
api_key = "your-api-key"
base_url = "https://your-endpoint.com/v1"
model = "your-model-name"

Then use it with:

klaw chat -p ollama
klaw chat -p lmstudio
klaw chat -p together

FAQ

"How is this different from OpenClaw?"

OpenClaw is powerful but complex — Node.js, multiple services, difficult deployment, hard to scale. klaw is a single Go binary with the same agent capabilities but kubectl-style operations. Deploy in seconds, scale by adding nodes.

"How is this different from LangChain/CrewAI?"

Those are frameworks for building agents. klaw is infrastructure for operating them. You could build agents with LangChain and deploy them on klaw.

"Is this about sandboxing agents?"

Partial. Namespaces provide logical isolation:

Scoped secrets — sales can't access support's API keys
Tool permissions — agents only get the tools you allow
Resource isolation — each namespace is independent

Important: Non-containerized agents have no filesystem sandboxing — they run under your user account and can access any file you can. For true process/filesystem isolation, run agents in Podman containers with klaw run.

"Can I run this on-prem?"

Yes. Single binary, no external dependencies. Run on your servers, your VPC, air-gapped environments. You control everything.

"What models are supported?"

300+ models via each::labs router, or direct connections to Anthropic, OpenAI, Google, Azure, Ollama, and any OpenAI-compatible endpoint.

Comparison

Feature	klaw	OpenClaw	LangChain	AutoGPT
Single Binary	✅	❌	❌	❌
kubectl-style CLI	✅	❌	❌	❌
Slack Control	✅	✅	Manual	❌
Namespaces	✅	❌	❌	❌
Built-in Cron	✅	❌	❌	❌
Distributed Mode	✅	❌	❌	❌
300+ Models	✅	✅	✅	❌
Easy Deployment	✅	❌	❌	❌
Enterprise Ready	✅	Partial	Partial	❌

Installation

Quick Install

curl -fsSL https://klaw.sh/install.sh | sh

From Source

git clone https://github.com/klawsh/klaw.sh.git
cd klaw && make build
sudo mv bin/klaw /usr/local/bin/

Verify

klaw version
# klaw v1.0.0 (darwin/arm64)

Contributing

git clone https://github.com/klawsh/klaw.sh.git
cd klaw
make build && make test
./bin/klaw chat

See CONTRIBUTING.md for guidelines.

Community

Documentation — Full docs and guides
GitHub Issues — Bug reports & feature requests
Discord — Chat with the community
Twitter — Updates & announcements

License

klaw is source-available under the each::labs License.

Free for internal business use, personal projects, and consulting. License required for multi-tenant SaaS or white-label distribution.

See LICENSE for details. Enterprise licensing: enterprise@eachlabs.ai

klaw — kubectl for AI Agents
Built by each::labs