Cohere Command A with Clanker Cloud — Open-Weights Enterprise AI for Infrastructure

Most teams evaluating AI for infrastructure management face the same tension: capability versus control. Managed API providers offer the most capable models, but data leaves your environment with every query. Open-source models keep data local, but typically sacrifice context length, tool use, and reliability.

Cohere Command A is open-weights, self-hostable, and designed for enterprise agentic tasks. Paired with Clanker Cloud, it can use Cohere's managed API or a customer-controlled model endpoint. A local endpoint can keep model prompts local; it does not make the public Clanker Cloud product fully air-gapped.

---

Cohere Command A — the enterprise-first model

Command A was released in March 2025. At 111 billion parameters, it is currently Cohere's most performant chat and agentic model. The API identifier is cohere.command-a-03-2025.

The model was not fine-tuned for enterprise use after the fact — it was designed for it from the outset. Cohere's positioning is explicit: "maximum performance with minimum hardware costs," with particular strength in business-critical agentic tasks, tool use, retrieval-augmented generation (RAG), and multilingual workloads across 10 or more languages.

Compared to Command R+, Command A delivers better throughput, stronger tool-use reliability, and more coherent coordination across multi-step agent workflows. These are not marginal improvements — they are the difference between an agent that completes a plan and one that stalls partway through.

The open-weights license means the model weights are downloadable. You can run Command A on your own GPU cluster, bare metal, or an isolated data center. That can keep model prompts and outputs inside the model environment; connected provider calls and the rest of the application data flow remain separate.

---

Why Command A is uniquely suited for enterprise infrastructure

256K context window

Infrastructure configuration does not fit neatly into small chunks. A non-trivial AWS environment might include hundreds of Terraform resources spread across dozens of files, CloudFormation stacks with cross-stack references, Kubernetes manifests for multiple namespaces, and IAM policies that interact with resources defined elsewhere. Chunking this input and feeding it to a model in pieces loses the cross-resource relationships that matter most for security audits, cost analysis, and dependency mapping.

Command A's 256K token context window holds an entire Terraform state file, a full CloudFormation template, and a suite of Kubernetes manifests simultaneously. Cross-resource relationships — an IAM role defined in one file affecting an EC2 instance defined in another — remain visible to the model in a single pass.

Among leading models, 256K is the largest available context window. Most peers sit at 128K or 200K. For large enterprise environments, that extra headroom is the difference between a complete audit and a partial one.

Open-weights and self-hosted deployment

Running Command A on your own infrastructure can keep model inference and prompts inside that environment. Organizations in financial services, healthcare, government, or defense contracting still need to assess cloud-provider calls, Clanker Cloud account services, disabled or enabled hosted features, and their applicable control requirements.

Command A is deployable on customer-controlled GPU infrastructure. With Clanker Cloud, raw provider credentials can stay in the normal desktop credential chain and the model endpoint can remain customer-controlled. Query results stay in that boundary only when the workflow is configured accordingly and hosted inference, sandbox, voice, and remote-control routes are not used.

Tool use and agent coordination

Command A was built with tool calling as a first-class capability. It handles multi-step agentic workflows without the degradation patterns common in models that were adapted for tool use after training. When Clanker Cloud routes a complex infrastructure task — spanning provider APIs, configuration files, and state data — Command A maintains plan coherence across the full sequence of tool calls.

---

Two deployment modes with Clanker Cloud

Mode 1: Cohere API (BYOK)

For teams that want Command A's capabilities without managing GPU infrastructure, the BYOK path uses Cohere's managed API:

1. Obtain an API key at dashboard.cohere.com
2. In Clanker Cloud, navigate to Settings → AI Model → Bring Your Own Key → Cohere
3. Paste the key and select cohere.command-a-03-2025

In direct desktop BYOK mode, the key is stored locally and used in requests sent directly from the machine to Cohere's API rather than through Clanker Cloud hosted inference. Cohere receives the key as the API credential plus the selected prompt and context under its terms. Account and other enabled hosted Clanker Cloud features remain separate routes.

For more on the BYOK model framework across all supported providers, see Clanker Cloud documentation.

Mode 2: Customer-controlled Command A endpoint

For customer-controlled model inference:

1. Download Command A weights from Cohere's model repository (Apache 2.0 license)
2. Run the model locally via Ollama or vLLM on your on-premises hardware
3. In Clanker Cloud Settings, configure the AI model endpoint to point to your local Ollama instance (e.g., http://localhost:11434)

In this configuration, model inference happens on customer-controlled hardware and no Cohere model API call is required. Cloud-provider calls still go to those providers, and ordinary Clanker Cloud account, security, download, or update traffic may occur. A fully isolated deployment requires a separately designed and verified environment; it is not created merely by selecting a local model endpoint.

This is one configuration enterprises can evaluate as part of a broader SOC 2, HIPAA, GDPR, or government control review. It does not create compliance or contractual approval by itself.

---

What Command A with 256K context can do

Full-stack infrastructure review

clanker ask "review my complete infrastructure configuration across all files and find security gaps, over-provisioning, and missing redundancy"

Feed the entire Terraform state, CloudFormation templates, and Kubernetes manifests in a single context. Command A tracks cross-resource dependencies — it understands that the IAM role in iam.tf grants permissions to the Lambda function in functions.tf, and that a permissive policy in one file has downstream effects across the stack.

This is the kind of holistic review that chunked-context approaches cannot reliably perform. A 128K model reviewing the same configuration in two passes will miss the relationship between resources that appear in different windows.

Multi-language team support

Command A supports 10 or more languages natively. Global enterprise teams can query their infrastructure in their working language without translation overhead or loss of precision:

clanker ask "zeige mir alle ungenutzten AWS-Ressourcen in der eu-west-1 Region"

This query in German — "show me all unused AWS resources in the eu-west-1 region" — returns the same structured analysis as the English equivalent. For distributed teams with engineers in Germany, Japan, France, or Spain, this removes a real friction point.

Agentic infrastructure workflows

clanker ask "plan and execute a cost optimization pass across my AWS account — identify savings, draft the Terraform changes, and prepare a summary for the team" --maker --apply

Command A's native tool use coordinates across multiple Clanker Cloud tools in sequence: querying provider APIs, analyzing the results, generating Terraform with the --maker flag, applying changes with --apply, and producing a formatted summary. The --agent-trace flag surfaces the full tool call sequence for review.

This is a multi-step agentic workflow that requires sustained plan coherence. Command A's design for enterprise agentic tasks makes it reliable here in ways that general-purpose models often are not.

---

Command A and Deep Research

Clanker Cloud's Deep Research feature fans out across every connected provider — AWS, GCP, Azure, Kubernetes, Cloudflare, and others — runs parallel analysis with multiple subagents, and returns prioritized findings organized by severity: cost drivers, misconfigurations, resilience gaps, and availability issues.

Command A is particularly well-suited as the backbone model for Deep Research runs in large environments:

clanker ask "run a deep research scan — prioritize security and compliance findings"

The 256K context window means the Deep Research agent can hold more infrastructure state simultaneously than most models allow. For enterprises with hundreds of resources across multiple regions and providers, this prevents the context truncation mid-audit that forces smaller-context models to drop earlier findings before completing a scan.

Deep Research results include severity levels, affected resources, evidence sources, estimated cost impact, and concrete action labels. Raw cloud-provider credentials stay in the normal desktop credential chain, while selected evidence follows the configured local, direct BYOK, or hosted inference route.

---

Command R7B — the lightweight Cohere option

For teams that do not need Command A's full capacity, Command R7B (released December 2024) offers a practical alternative for frequent, lightweight queries.

At 7 billion parameters, Command R7B runs on a single RTX 3090 or an Apple M-series Mac. It is optimized for RAG and tool use within a smaller footprint:

clanker ask "quick check — any new alerts or failures in the last 15 minutes"

Use cases for Command R7B include lightweight monitoring agents, high-frequency small queries where latency matters more than depth, and edge deployments. It can be deployed as a sidecar on Kubernetes nodes to provide local infrastructure reasoning without egress — a useful pattern for clusters where network egress is restricted or expensive.

Command A and Command R7B serve different points on the capability-resource tradeoff. Many teams will use both: Command R7B for continuous low-overhead monitoring, Command A for periodic deep audits and complex agentic workflows.

---

Compliance and data residency

Self-hosted Command A can reduce model-provider data transfer, but neither the model choice nor Clanker Cloud's local credential path establishes SOC 2, HIPAA, GDPR, government, or residency compliance.

The data flow in a customer-controlled model configuration:

• Raw infrastructure credentials: held in the normal local desktop credential chain and not sent to Clanker Cloud
• Model inference: runs on your hardware, on your network
• Query and model logs: can remain on-prem when configured locally and hosted features are disabled
• Provider API calls: go directly from your machine to AWS/GCP/Azure — no proxy, no intermediary

Organizations subject to data residency requirements must still document all provider calls, Clanker Cloud account and security traffic, support access, logging, retention, updates, and every enabled hosted feature. Hosted personal-data use requires an executed DPA; hosted PHI or government use requires the applicable signed terms and a verified active protected environment.

The AI DevOps for Teams page covers how Clanker Cloud handles team deployments in regulated environments, including credential scoping and audit trail configuration.

---

Command A via MCP for enterprise agents

Command A's tool-use capability extends beyond direct CLI interaction. Teams building enterprise automation agents — deployment pipelines, incident response workflows, cost governance systems — can connect those agents to Clanker Cloud's infrastructure tooling via MCP:

clanker mcp --transport http --listen 127.0.0.1:39393

With the MCP server running, any agent built on Command A (via Cohere API or self-hosted) can call Clanker Cloud infrastructure tools using the standard MCP protocol. The primary tool for natural-language infrastructure queries is clanker_route_question.

A practical pattern: an enterprise deployment agent that queries infrastructure state before every release. Before applying a Terraform plan, the agent calls clanker_route_question to confirm the target environment is healthy and that the planned changes do not conflict with current state. This happens programmatically, without human intervention, using Command A as the reasoning layer.

For detailed MCP integration guidance, see Clanker Cloud for AI Agents. Teams building production agent workflows should also review the vibe-coding-to-production guide for patterns that move agent-assisted work safely through staging and into production.

---

FAQ

What is Cohere Command A and can I self-host it?

Command A is Cohere's flagship enterprise model, released March 2025. It has 111 billion parameters, a 256K token context window, and is available under an Apache 2.0 compatible open-weights license. You can download the model weights and run Command A on your own GPU hardware — 2x A100 80GB, or 4x RTX 4090 with AWQ quantization — with no dependency on Cohere's API. The model is designed for agentic tasks, tool use, RAG, and multilingual workloads.

How do I use Cohere Command A with Clanker Cloud?

There are two paths. For the managed API route, obtain a Cohere API key at dashboard.cohere.com, then go to Clanker Cloud Settings → AI Model → Bring Your Own Key → Cohere and paste the key. Select the cohere.command-a-03-2025 model identifier. For self-hosted inference, run Command A via Ollama or vLLM on your local hardware and point Clanker Cloud's model endpoint to your local Ollama instance. Full setup documentation is at docs.clankercloud.ai.

Why does the 256K context window matter for infrastructure management?

Infrastructure configurations involve cross-resource dependencies that break when the context is chunked. An IAM policy in one file affects resources defined in another; a VPC configuration constrains subnets and security groups throughout the stack. A model with a 128K or 200K context window reviewing a large environment in multiple passes loses those cross-file relationships. Command A's 256K window holds an entire multi-file infrastructure configuration in a single context, enabling accurate dependency analysis, security audits, and cost reviews without truncation.

Is Cohere Command A suitable for air-gapped enterprise environments?

Command A can run within your network perimeter, and a local endpoint avoids sending prompts to Cohere's API. The normal Clanker Cloud desktop workflow can keep raw provider credentials local, but the public product still has account, security, download, update, and optional hosted-service paths. Do not call the result air-gapped or approved for financial, healthcare, government, or defense data unless a separately delivered environment has been verified against the required controls and contracts.

---

Get started

Clanker Cloud is in public beta with no cost on the Beta tier. The demo walks through provider connection, model configuration, and your first infrastructure query in under five minutes.

To create an account and connect your first provider, go to clankercloud.ai/account. For teams evaluating the self-hosted Command A path, start with the documentation — it covers Ollama endpoint configuration, MCP setup, and team credential scoping.

Questions about deployment options and compliance configurations are covered in the FAQ.