--- sidebar_position: 3 title: LLM Providers --- # 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. :::tip 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](./configuration.md) 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): ```bash 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](https://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-ai} `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. ```bash export PENTESTSWARM_ORCHESTRATOR_API_KEY=your-together-key pentestswarm scan --scope --provider together --swarm ``` Get a key at [api.together.xyz](https://api.together.xyz). :::note Explicit model wins If you set an explicit **per-agent model** in your [config](./configuration.md), 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) ```bash export PENTESTSWARM_ORCHESTRATOR_API_KEY=sk-ant-... pentestswarm scan --scope --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. ```yaml 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 ```bash export PENTESTSWARM_ORCHESTRATOR_API_KEY=AIza... pentestswarm scan --scope --provider gemini --swarm ``` Large context window and a free tier — get a key from [Google AI Studio](https://aistudio.google.com/apikey). ### 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) ```yaml 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 ``` ```bash pentestswarm scan --scope --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. ```bash export PENTESTSWARM_ORCHESTRATOR_API_KEY=sk-orca-... pentestswarm scan --scope --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 ```bash # install Ollama and pull a tool-calling model, then: pentestswarm scan --scope --provider ollama --swarm ``` ### LM Studio Load a model in LM Studio, enable its local server, then: ```bash pentestswarm scan --scope --provider lmstudio --swarm ``` :::note 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 ` — 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](./cost-and-safety.md). To pin models, keys, or override individual agents, see [Configuration](./configuration.md).