@waspdev95/anyagent

Claude Code, Codex, OpenCode and nine others — on OpenRouter, DeepSeek, Groq, a local Ollama, or anywhere else. Each of them has its own way of pointing somewhere else. This is one command for all of them.

npm install -g @waspdev95/anyagent
anyagent

  anyagent - pick one, and I will ask for a key and a model

❯ claude    Claude Code       ✔ deepseek/deepseek-v4-pro
  codex     Codex             ✔ openai/gpt-5.6-sol
  opencode  OpenCode          ✔ moonshotai/kimi-k3
  droid     Droid               not installed

  model     Choose the AI model    openrouter / deepseek-v4-pro
  key       Add an API key         1 saved
  check     Check my setup
  quit      Exit

Pick one and it runs. It asks for a key the first time, then never again.

That is the whole thing. Everything below is optional.

Typing instead

anyagent claude                        # run it
anyagent codex -m openai/gpt-5.6-sol   # different model, just this once
anyagent model                         # change the model for everything
anyagent key openrouter                # save a key
anyagent ls                            # what you have

Flags anyagent does not recognise go straight to the agent, including --help: anyagent claude --resume --dangerously-skip-permissions. anyagent's own flags go before the name — anyagent --json claude --dry-run.

Already have OPENROUTER_API_KEY (or GROQ_API_KEY, DEEPSEEK_API_KEY, …) exported? Then there is nothing to set up — anyagent uses it.

Instead of doing this by hand

Every guide for pointing an agent somewhere else reads like this. It is OpenRouter's, for Claude Code:

export ANTHROPIC_BASE_URL="https://openrouter.ai/api"
export ANTHROPIC_AUTH_TOKEN="$OPENROUTER_API_KEY"
export ANTHROPIC_API_KEY=""              # must be empty, or Claude Code
                                         # authenticates against Anthropic instead
export ANTHROPIC_DEFAULT_OPUS_MODEL="..."
export ANTHROPIC_DEFAULT_SONNET_MODEL="..."
export ANTHROPIC_DEFAULT_HAIKU_MODEL="..."
export CLAUDE_CODE_SUBAGENT_MODEL="..."

Then the same problem again, solved differently, for every other agent: Codex needs a TOML profile with wire_api = "responses", OpenCode wants JSON, Droid wants an entry in ~/.factory/settings.json, DeepSeek Harness wants a YAML patch.

anyagent claude    # this

Two things anyagent does that the manual route does not:

• Nothing is exported into your shell. The variables exist for the launched process only, so a plain claude still uses your Anthropic account, and two terminals can run two models at once.
• The real context window is passed through. Claude Code assumes 200K for a model it does not recognise, so a 1M-token model compacts far too early and a small one too late. anyagent reads the actual limit from the catalog and sets CLAUDE_CODE_MAX_CONTEXT_TOKENS.

Common pairings

Claude Code on DeepSeek	anyagent claude -m deepseek/deepseek-v4-pro
Claude Code on Kimi	anyagent claude -m moonshotai/kimi-k3
Claude Code on GLM / Z.ai	anyagent claude --provider zai
Claude Code on a local Ollama	anyagent claude --provider ollama -m qwen3-coder
Codex on GPT-5.6	anyagent codex -m openai/gpt-5.6-sol
Codex on Groq	anyagent codex --provider groq
OpenCode on anything	anyagent opencode -m <model>

Claude Code needs an Anthropic-compatible endpoint and Codex needs OpenAI's Responses API, so not every pairing exists. anyagent compat says which do, before you start rather than three turns in.

What it works with

Agents: claude · codex · opencode · copilot · droid · dsh · pi · openclaw · hermes · cline · qwen · pool

Providers: 183 with 7,000+ models, from the models.dev catalog — OpenRouter, DeepSeek, Z.ai, Moonshot, Groq, Cerebras, Together, xAI, OpenAI, Anthropic and the rest. Local Ollama, LM Studio, llama.cpp and vLLM need no key at all.

Anything else works with an explicit endpoint:

anyagent opencode --base-url http://localhost:8080/v1 -m my-model
anyagent exec -- aider --model deepseek/deepseek-v4-pro  # tools with no integration

Your files stay yours

• Seven of the twelve agents are never written to — the settings live in the launched process, so anyagent claude and a plain claude can run at the same time on different models.
• When a file must change, the original is backed up and anyagent restore puts it back.
• Keys sit in ~/.anyagent/credentials.json (mode 0600), or your OS keychain. They are never printed in full and never appear on a command line.

More

anyagent help --all	every command
Agents · Providers	the full lists
Troubleshooting	when something is off
Architecture · Security	how it works, where keys go
Contributing	a new agent is about forty lines

MIT licensed.