Models

Decepticon routes every LLM call through a LiteLLM proxy that abstracts Anthropic, OpenAI, Google, MiniMax, DeepSeek, xAI, Mistral, OpenRouter, Nvidia NIM, local Ollama, self-hosted vLLM, plus six subscription OAuth handlers (Claude Code / ChatGPT / Gemini Advanced / Copilot Pro / SuperGrok / Perplexity Pro) behind a single endpoint. The model assigned to each agent — and the model that takes over when the primary fails — is computed at startup from your credentials inventory plus the active profile.

You don't pick agent-by-agent models manually. You tell Decepticon which credentials you have, in what order of preference; it builds the chain.

---

How model selection works

Three orthogonal axes:

Axis	Values	Decided by
Tier	HIGH / MID / LOW	Agent (e.g. orchestrator → HIGH, recon → LOW), overridable by profile
AuthMethod	API: anthropic_api / openai_api / google_api / minimax_api / deepseek_api / xai_api / mistral_api / openrouter_api / nvidia_api OAuth: anthropic_oauth / openai_oauth / google_oauth / copilot_oauth / grok_oauth / perplexity_oauth Local: ollama_local	Your credentials inventory
Profile	eco / max / test	DECEPTICON_MODEL_PROFILE

For each agent, Decepticon resolves a tier (from the profile) and walks your AuthMethod priority list, emitting the model identifier that method provides at that tier. The first hit is the primary; every remaining hit is queued as a fallback in priority order. langchain's ModelFallbackMiddleware walks the queue on primary failure, trying each method in turn until one succeeds.

Tier × AuthMethod matrix

	HIGH	MID	LOW
anthropic_api	anthropic/claude-opus-4-8	anthropic/claude-sonnet-5	anthropic/claude-haiku-4-5
anthropic_oauth	auth/claude-opus-4-8	auth/claude-sonnet-5	auth/claude-haiku-4-5
openai_api	openai/gpt-5.5	openai/gpt-5.4	openai/gpt-5-nano
openai_oauth	auth/gpt-5.5	auth/gpt-5.4	auth/gpt-5.4-mini
google_api	gemini/gemini-2.5-pro	gemini/gemini-2.5-flash	gemini/gemini-2.5-flash-lite
google_oauth	gemini-sub/gemini-2.5-pro	gemini-sub/gemini-2.5-flash	— (falls through)
minimax_api	minimax/MiniMax-M3	minimax/MiniMax-M2.7-highspeed	— (falls through)
deepseek_api	deepseek/deepseek-v4-pro	deepseek/deepseek-v4-flash	deepseek/deepseek-v4-flash
xai_api	xai/grok-4.3	xai/grok-4-1-fast-reasoning	— (falls through)
grok_oauth	grok-sub/grok-4.3	grok-sub/grok-4-1-fast-reasoning	— (falls through)
mistral_api	mistral/mistral-large-latest	mistral/codestral-latest	— (falls through)
openrouter_api	openrouter/anthropic/claude-opus-4-8	openrouter/anthropic/claude-sonnet-4-6	openrouter/anthropic/claude-haiku-4-5
nvidia_api	nvidia_nim/meta/llama-3.3-70b-instruct	nvidia_nim/nvidia/llama-3.1-nemotron-70b-instruct	nvidia_nim/meta/llama-3.2-3b-instruct
copilot_oauth	copilot/gpt-5.5	copilot/claude-sonnet-4-6	copilot/gpt-5.4-mini
perplexity_oauth	pplx-sub/sonar-pro	pplx-sub/sonar	— (falls through)
ollama_local	ollama_chat/<OLLAMA_MODEL>	ollama_chat/<OLLAMA_MODEL>	ollama_chat/<OLLAMA_MODEL>

ollama_local collapses across tiers — local GPUs typically run a single model — and the slug is whatever you pulled (e.g. qwen3-coder:30b). When a method has no model at the requested tier (MiniMax LOW, Mistral LOW, ...), the resolver skips it and continues with the next method in your priority list.

The matrix above lists the primary tier-mapped providers. Decepticon resolves the same tier × AuthMethod chain for additional providers not shown above: managed cloud platforms (AWS Bedrock, GCP Vertex AI, Azure OpenAI), further direct APIs (Groq, Together, Fireworks, Cohere, Moonshot, Z.ai, DashScope, Cerebras, Xiaomi MiMo, Baidu Qianfan), multi-vendor gateways (GitHub Models, Hugging Face, OpenCode, Vercel AI Gateway, Cloudflare AI Gateway, Venice, NanoGPT, Synthetic, ZenMux), and local / self-hosted OpenAI-compatible servers (vLLM, Ollama Cloud, LM Studio, llama.cpp, and custom endpoints). AuthMethod and METHOD_MODELS in packages/decepticon-core/decepticon_core/types/llm.py are the authoritative source for the full provider list and each provider's tier→model mapping.

---

Profiles

DECEPTICON_MODEL_PROFILE (default: eco) controls which tier each agent runs at.

eco — per-agent tier (production default)

Each agent runs at the tier suited to its workload:

Tier	Agents
HIGH	decepticon, exploit, exploiter, patcher, contract_auditor, analyst, vulnresearch
MID	detector, verifier, blue_cell, postexploit, ad_operator, cloud_hunter, reverser, phisher, mobile_operator, iot_operator, ics_operator, forensicator, supply_chain_operator
LOW	soundwave, recon, scanner, wireless_operator, osint_operator

max — every agent on HIGH

For high-value targets where accuracy outweighs cost. Forces every agent to the HIGH tier.

test — every agent on LOW

For development / CI. Forces every agent to the cheapest tier (Haiku-class).

---

Credentials inventory

Your inventory is built at startup from environment variables, written by decepticon onboard.

# Priority order (first = preferred). When unset, the factory's built-in
# `_DEFAULT_AUTH_PRIORITY` order applies (OAuth subscriptions ahead of their
# paid per-token peers, hosted providers before local endpoints, with
# unconfigured methods skipped at runtime). See `_DEFAULT_AUTH_PRIORITY` in
# `decepticon/llm/factory.py` for the full ordered list.
DECEPTICON_AUTH_PRIORITY=anthropic_oauth,openai_api

# Set true if you have an active Claude Code OAuth subscription
# (anthropic_oauth in the priority list above).
DECEPTICON_AUTH_CLAUDE_CODE=true

# Per-method credentials. Placeholder values (`your-..-key-here`) are
# treated as "not configured" and silently dropped from the inventory.
ANTHROPIC_API_KEY=sk-ant-...
OPENAI_API_KEY=sk-...
GEMINI_API_KEY=AIza...
MINIMAX_API_KEY=eyJ...
DEEPSEEK_API_KEY=sk-...
XAI_API_KEY=xai-...
MISTRAL_API_KEY=...
OPENROUTER_API_KEY=sk-or-...
NVIDIA_API_KEY=nvapi-...

# Local LLM (no API key — point at your Ollama server)
OLLAMA_API_BASE=http://host.docker.internal:11434
OLLAMA_MODEL=qwen3-coder:30b

The factory walks the priority list, drops methods whose detection check fails (placeholder API key, or DECEPTICON_AUTH_CLAUDE_CODE=false), and uses what's left.

---

Fallback chain examples

All examples assume the eco profile.

Single API key (Anthropic only)

DECEPTICON_AUTH_PRIORITY=anthropic_api
ANTHROPIC_API_KEY=sk-ant-...

Agent (tier)	Primary	Fallback
decepticon (HIGH)	anthropic/claude-opus-4-7	—
detector (MID)	anthropic/claude-sonnet-4-6	—
recon (LOW)	anthropic/claude-haiku-4-5	—

No fallback — only one credential.

Single API key (OpenAI only)

DECEPTICON_AUTH_PRIORITY=openai_api
OPENAI_API_KEY=sk-...

Agent (tier)	Primary	Fallback
decepticon (HIGH)	openai/gpt-5.5	—
detector (MID)	openai/gpt-5.4	—
recon (LOW)	openai/gpt-5-nano	—

Claude Code OAuth + Anthropic API (subscription primary, paid fallback)

DECEPTICON_AUTH_PRIORITY=anthropic_oauth,anthropic_api
DECEPTICON_AUTH_CLAUDE_CODE=true
ANTHROPIC_API_KEY=sk-ant-...

Agent (tier)	Primary	Fallback
decepticon (HIGH)	auth/claude-opus-4-7	anthropic/claude-opus-4-7
detector (MID)	auth/claude-sonnet-4-6	anthropic/claude-sonnet-4-6
recon (LOW)	auth/claude-haiku-4-5	anthropic/claude-haiku-4-5

OAuth runs primary (no API cost). When the subscription quota hits, fallback drops to the paid API key — same model family, same quality.

Mixed providers (Anthropic + OpenAI)

DECEPTICON_AUTH_PRIORITY=anthropic_api,openai_api
ANTHROPIC_API_KEY=sk-ant-...
OPENAI_API_KEY=sk-...

Agent (tier)	Primary	Fallback
decepticon (HIGH)	anthropic/claude-opus-4-7	openai/gpt-5.5
detector (MID)	anthropic/claude-sonnet-4-6	openai/gpt-5.4
recon (LOW)	anthropic/claude-haiku-4-5	openai/gpt-5-nano

Cross-provider fallback — when Anthropic hits a rate limit or outage, OpenAI takes over seamlessly.

Local Ollama only (offline / cost-free)

DECEPTICON_AUTH_PRIORITY=ollama_local
OLLAMA_API_BASE=http://host.docker.internal:11434
OLLAMA_MODEL=qwen3-coder:30b

Agent (tier)	Primary	Fallback
decepticon (HIGH)	ollama_chat/qwen3-coder:30b	—
detector (MID)	ollama_chat/qwen3-coder:30b	—
recon (LOW)	ollama_chat/qwen3-coder:30b	—

Same model across all tiers because local hardware typically can't run three different models simultaneously. The ollama_chat/ provider routes to Ollama's /api/chat endpoint, the only one that supports tool/function calling — the legacy ollama/ provider hits /api/generate and is rejected at LiteLLM-config-merge time because Decepticon agents always emit tool calls.

Two probes guard the wiring end-to-end:

1. decepticon onboard — the wizard hits /api/tags and /api/show on the host, filters to models that report tools in their capabilities, and presents only that list as the OLLAMA_MODEL choice. If Ollama is unreachable or has no tool-capable models pulled, the wizard refuses to write .env and prints the exact remediation steps.

2. litellm container startup — re-runs the same checks from inside the container, the only place that can detect OLLAMA_HOST=127.0.0.1-only bindings (which look fine to the wizard's host probe but are invisible from the container). Every diagnostic appears in decepticon logs litellm prefixed with [decepticon ollama].

Local Ollama + cloud fallback

DECEPTICON_AUTH_PRIORITY=ollama_local,anthropic_api
OLLAMA_API_BASE=http://host.docker.internal:11434
OLLAMA_MODEL=qwen3-coder:30b
ANTHROPIC_API_KEY=sk-ant-...

Agent (tier)	Primary	Fallback
decepticon (HIGH)	ollama_chat/qwen3-coder:30b	anthropic/claude-opus-4-7
detector (MID)	ollama_chat/qwen3-coder:30b	anthropic/claude-sonnet-4-6
recon (LOW)	ollama_chat/qwen3-coder:30b	anthropic/claude-haiku-4-5

Local model handles routine work; when the local model fails (OOM, context overflow, hardware fault), Anthropic takes over for that request only.

Local llama.cpp (GGUF) only

DECEPTICON_AUTH_PRIORITY=llamacpp_local
LLAMACPP_API_BASE=http://host.docker.internal:8080/v1
LLAMACPP_MODEL=qwen2.5-coder-7b-instruct-q4_k_m

Agent (tier)	Primary	Fallback
decepticon (HIGH)	llamacpp/qwen2.5-coder-7b-instruct-q4_k_m	—
detector (MID)	llamacpp/qwen2.5-coder-7b-instruct-q4_k_m	—
recon (LOW)	llamacpp/qwen2.5-coder-7b-instruct-q4_k_m	—

Run a GGUF model with llama.cpp's OpenAI-compatible server:

llama-server -m /path/to/qwen2.5-coder-7b-instruct-q4_k_m.gguf --port 8080
# server listens on http://localhost:8080/v1

Like Ollama, the model collapses across tiers — llama-server runs one GGUF at a time. The llamacpp/<model> route remaps to LiteLLM's openai/ provider (since llama.cpp does not have a dedicated LiteLLM provider) plus a custom api_base pointed at LLAMACPP_API_BASE. This is the integration the maintainer recommended on issue #151 — it reuses the existing LiteLLM provider machinery and avoids vendoring llama-cpp-python as a runtime dependency.

Tool calling works as long as the GGUF was trained or fine-tuned with function-calling support (Qwen 2.5 Coder, Llama 3.1 Instruct, Mistral Small Instruct all qualify). The llama-server request format is identical to OpenAI's; LiteLLM translates the agent's tool calls through unchanged.

MiniMax-only (LOW gap)

DECEPTICON_AUTH_PRIORITY=minimax_api
MINIMAX_API_KEY=eyJ...

Agent (tier)	Primary	Notes
decepticon (HIGH)	minimax/MiniMax-M3	OK
detector (MID)	minimax/MiniMax-M2.7-highspeed	OK
recon (LOW)	(role unassigned)	MiniMax has no LOW model and no fallback method

The Recon/Scanner/Soundwave roles fail to initialize. Add a second AuthMethod (e.g. openai_api) to fill the LOW slot.

---

Failover behavior

ModelFallbackMiddleware (from langchain.agents.middleware) watches every LLM call. On primary failure (provider outage, 429 rate limit, context overflow, network error), it transparently retries each queued fallback in order until one succeeds. Agents see no interruption — same conversation history, same tool call.

The middleware receives the full chain [primary, *fallbacks] from LLMFactory.get_fallback_models(role). If the user has all five AuthMethods configured, that's a five-deep chain; with a single credential it's primary-only and the middleware short-circuits. The chain length scales with credentials inventory — no upper cap, no silent truncation. Only when every method fails does the agent surface the error.

---

LiteLLM proxy

All traffic flows through the LiteLLM container on port 4000. The proxy provides:

• Unified API — agents call one endpoint, model identifier picks the backend
• Usage tracking — tokens per model per agent role
• Rate limiting — per-provider knobs
• Cost attribution — billing data aggregated across providers

Configuration: config/litellm.yaml. Authentication: LITELLM_MASTER_KEY in .env.

---

Adding a model

To wire in a new provider model:

1. Add a model_list entry to config/litellm.yaml with the LiteLLM provider/model identifier and the env var that holds the key.
2. Add the model identifier to the appropriate cell of METHOD_MODELS in decepticon_core.types.llm.
3. If introducing a new AuthMethod, also add it to AuthMethod, the factory's _API_METHOD_ENV map, and the onboard wizard's option list.

Tests in tests/unit/llm/test_models.py will catch dropped tiers or missing matrix entries.

Dynamic model registration (Ollama, custom gateways, ad-hoc overrides)

You don't need to edit YAML to add an Ollama model. Set the env vars below and litellm_dynamic_config.py registers the route at proxy startup:

Env var	What it does
OLLAMA_MODEL=<tag> + OLLAMA_API_BASE=<url>	Registers ollama_chat/<tag> automatically. Used by the ollama_local AuthMethod.
DECEPTICON_MODEL=<provider/model>	Registers a global override (e.g. groq/llama-3.3-70b-versatile).
DECEPTICON_MODEL_<ROLE>=<provider/model>	Per-role override (e.g. DECEPTICON_MODEL_RECON=ollama_chat/llama3.2).
DECEPTICON_LITELLM_MODELS=<a,b,c>	Bulk register multiple ids without editing YAML.
CUSTOM_OPENAI_API_BASE + CUSTOM_OPENAI_API_KEY	OpenAI-compatible gateway. Use custom/<model> in the override env.

The proxy logs [decepticon] registered N dynamic model route(s) at startup so you can confirm what got picked up.

---

Subscription OAuth Providers

Use monthly subscriptions instead of per-token API billing. All providers use custom LiteLLM handlers that authenticate via OAuth/session tokens. ChatGPT in particular reads from the Codex CLI credential store (~/.codex/auth.json), so a host-side codex login flows into the running container without a rebuild.

Subscription	AuthMethod	Models	Handler
Claude Max/Pro/Team	anthropic_oauth	auth/claude-opus, sonnet, haiku	claude_code_handler.py
ChatGPT Pro/Plus/Team	openai_oauth	auth/gpt-5.5, gpt-5.4, gpt-5.4-mini (+ gpt-5.3-codex for code roles)	codex_chatgpt_handler.py (reads ~/.codex/auth.json)
Gemini Advanced	google_oauth	gemini-sub/gemini-2.5-pro, flash	gemini_handler.py
Copilot Pro	copilot_oauth	copilot/gpt-5.5, claude-sonnet-4-6, gpt-5.4-mini (+ gpt-5.3-codex)	copilot_handler.py
SuperGrok	grok_oauth	grok-sub/grok-4.3, grok-4-1-fast-reasoning	grok_handler.py
Perplexity Pro	perplexity_oauth	pplx-sub/sonar-pro, sonar	perplexity_handler.py

Enable in .env:

DECEPTICON_AUTH_CLAUDE_CODE=true     # Claude subscription
DECEPTICON_AUTH_CHATGPT=true         # ChatGPT subscription
DECEPTICON_AUTH_GEMINI=true          # Gemini Advanced
DECEPTICON_AUTH_COPILOT=true         # Copilot Pro
DECEPTICON_AUTH_GROK=true            # SuperGrok
DECEPTICON_AUTH_PERPLEXITY=true      # Perplexity Pro

For full setup instructions including token extraction, see Setup Guide.