LLM Providers

Pentest Swarm AI is the harness, not the model. Point it at whichever LLM you want — a frontier cloud model, an open cyber-benchmark leader, or a fully local one.

One key, whole swarm

All agents inherit from a single provider config. Set one provider and one key, and the entire swarm — orchestrator plus every agent — uses it. There is no per-agent key wrangling. (You can override individual agents in Configuration if you want to — but you never have to.)

At a glance

Provider	Key	Cost	Shape
together	one key	~$1 / run baseline	Multi-model — routes several open models by task
ollama / lmstudio	none	free, air-gapped	Local, one model
claude / openai / gemini	one key	per vendor	Single model — every agent shares it
bedrock	AWS profile or keys	per model	Single model — any Bedrock Converse model you set with --model
orcarouter	one key	per gateway	Gateway fronting many frontier models

Every provider must support native tool / function calling — that's what lets the swarm operate tools instead of just describing them.

Setting the key

For any cloud provider, the API key is read from an environment variable (or set in the launcher UI / config file):

export PENTESTSWARM_ORCHESTRATOR_API_KEY=your-key-here

Select the provider with the --provider flag on scan, in pentestswarm run, or in your config's orchestrator.provider.

Supported providers

Provider	--provider	Key needed	Notes
Together AI	together	yes	Hosted Llama / Qwen / DeepSeek / GLM; key from api.together.xyz
Claude	claude	yes	Anthropic; best quality, prompt caching
OpenAI	openai	yes	OpenAI and any OpenAI-compatible endpoint
Gemini	gemini	yes	Google; large context, free tier available
OrcaRouter	orcarouter	yes	One gateway for Claude / GPT + other frontier models
AWS Bedrock	bedrock	no (AWS credentials)	Any Converse model ID or inference profile; set it with --model
Ollama	ollama	no	100% local, air-gapped
LM Studio	lmstudio	no	100% local, GUI model management

Together AI — first-class multi-model mode

together is the first-class multi-model provider. Instead of one model doing every job, the swarm routes several open models by task — each role gets the model that's best (and best-priced) for it:

Task	Model	Approx. price
Recon + report	Llama-3.3-70B	~$0.88 / Mtok
Classifier + exploit	GLM-5.3-Flash	$0.15 in / $0.50 out / Mtok

The reasoning-heavy roles run on Z.ai's GLM-5.3-Flash — the top-ranked open-weight model on the agentic leaderboard (Sept 2026), with near-frontier exploitation reasoning. It's both stronger and cheaper than the DeepSeek-V3 / Qwen2.5 it replaced, so the swarm got sharper and its per-run cost dropped.

You supply one key and the endpoint is auto-configured — no per-model setup. Because the cheap fast model carries the high-volume recon/report work and GLM-5.3-Flash is reserved for the reasoning-heavy classify/exploit steps, routing keeps a typical run near (or under) the ~$1 single-model baseline rather than multiplying cost.

export PENTESTSWARM_ORCHESTRATOR_API_KEY=your-together-key
pentestswarm scan <target> --scope <target> --provider together --swarm

Get a key at api.together.xyz.

Explicit model wins

If you set an explicit per-agent model in your config, that override always wins over the automatic routing for that agent. The routing only applies to agents you leave unset.

Under the hood Together speaks the OpenAI Chat-Completions API, so you can also drive it via the generic openai provider with the endpoint https://api.together.xyz/v1 — but you'd lose the automatic multi-model routing, so prefer --provider together.

Claude (Anthropic)

export PENTESTSWARM_ORCHESTRATOR_API_KEY=sk-ant-...
pentestswarm scan <target> --scope <target> --provider claude --swarm

Best out-of-the-box quality, and prompt caching is enabled by default for the recon and classifier agents to cut cost and latency.

OpenAI (and any OpenAI-compatible endpoint)

The openai provider works with OpenAI itself and any service that exposes a Chat-Completions /v1 API — DeepSeek, Groq, Together, and more. The only hard requirement is native tool / function calling. Point it at the vendor's base URL in your config (orchestrator.endpoint) and set the key.

orchestrator:
  provider: "openai"
  endpoint: "https://api.together.xyz/v1"
  model: "zai-org/GLM-5.3"     # or Qwen/..., deepseek-ai/..., etc.
  api_key: ""                   # or export PENTESTSWARM_ORCHESTRATOR_API_KEY
  context_window: 128000

Gemini

export PENTESTSWARM_ORCHESTRATOR_API_KEY=AIza...
pentestswarm scan <target> --scope <target> --provider gemini --swarm

Large context window and a free tier — get a key from Google AI Studio.

AWS Bedrock

One model for every agent, invoked through Bedrock's Converse API, so the same config reaches GLM, Claude, Llama, Nova, and any other model the account can call. The model is required — pass --model or set orchestrator.model to a foundation-model ID or an inference profile (for example zai.glm-5).

Bedrock does not take api_key. Credentials resolve in this order:

1. aws_access_key_id + aws_secret_access_key (and optional aws_session_token)
2. aws_profile (shared AWS config)
3. The default AWS chain (AWS_PROFILE, AWS_ACCESS_KEY_ID, SSO, instance role)

orchestrator:
  provider: "bedrock"
  model: "zai.glm-5"          # required; or e.g. us.anthropic.claude-sonnet-4-5-20250929-v1:0
  region: "us-east-1"
  aws_profile: "rohit"        # or set access keys / rely on AWS_PROFILE

pentestswarm scan <target> --scope <target> --swarm \
  --provider bedrock --model zai.glm-5 \
  --aws-profile rohit --aws-region us-east-1

Access keys belong in config or PENTESTSWARM_ORCHESTRATOR_AWS_ACCESS_KEY_ID / PENTESTSWARM_ORCHESTRATOR_AWS_SECRET_ACCESS_KEY, not on the command line. The model must support Converse tool use so the swarm can operate tools.

OrcaRouter

A single gateway endpoint that fronts Claude, GPT, and other frontier models with gateway-level security.

export PENTESTSWARM_ORCHESTRATOR_API_KEY=sk-orca-...
pentestswarm scan <target> --scope <target> --provider orcarouter --swarm

Local providers (no key, no cloud)

For full privacy and zero API cost, run the swarm entirely on your own machine. No API key is required for either of these.

Ollama

# install Ollama and pull a tool-calling model, then:
pentestswarm scan <target> --scope <target> --provider ollama --swarm

LM Studio

Load a model in LM Studio, enable its local server, then:

pentestswarm scan <target> --scope <target> --provider lmstudio --swarm

Tool calling required

Whatever model you choose — local or hosted — must support native tool / function calling. That's what lets the swarm actually operate tools rather than just describe them.

Keeping cost predictable

On a paid provider, cap what a run can spend with --budget <usd> — a hard per-run USD ceiling that winds the campaign down gracefully (and still writes the report) the moment it's reached. Local providers have no cost, so the cap is n/a. See Cost & Safety.

To pin models, keys, or override individual agents, see Configuration.