@inference/fast

Connect Claude Code, Codex, Grok, OpenCode, and Pi to fast models on Inference.net. The CLI signs in through the Inference dashboard, lets you choose a team and project when more than one is available, and creates one machine-scoped project API key.

Install

npm install -g @inference/fast

Use

fast login
fast models
fast claude on
fast claude on --model glm-5.2
fast claude on --main glm-5.2 --sonnet kimi-k3-fast
fast codex on --model glm-5.2
fast grok on --model glm-5.2
fast opencode on --build kimi-k3-fast --plan deepseek-v4-flash-0731
fast opencode status
fast pi off
fast claude on --new-key
fast key rotate
fast status
fast logout

fast models lists every model in the catalog with its context window, price per million tokens, and capabilities; pass a substring to filter (fast models kimi), --endpoint messages to see only models an agent can talk to, or --json for scripts.

Calling a coding agent without on is equivalent to on, so fast claude is enough. Claude Code has a default model per slot (main, opus, sonnet, and fable use the default model; haiku and subagents use deepseek-v4-flash-0731). --model sets every slot at once; --main, --opus, --sonnet, --haiku, --fable, and --subagents override one slot and win over --model. OpenCode accepts --build and --plan the same way for its two primary agents. Every on registers every model from the catalog in the agent's own model picker and snapshots the agent's original config before changing it. off restores that snapshot byte-for-byte. Every on re-validates the saved machine key against the gateway. --new-key (or fast key rotate) archives that key, creates a new one, and rewrites every connected agent with it. logout restores every connected agent and disables the API key created for this machine. The inference gateway can retain an already-used key in its authentication cache for up to three minutes before it starts returning unauthorized responses.

Files

• Owner-private CLI state and transport credentials: ~/.inference-fast
• Claude Code: ~/.claude/settings.json
• Codex: ~/.codex/config.toml
• OpenCode: ~/.config/opencode/opencode.json
• Pi: ~/.pi/agent/{settings,models,auth}.json

All CLI and harness files that can contain a credential are written with owner- only permissions. Harness backups live under ~/.inference-fast/backups.

Local development

With relay, the inference gateway, and fast-web running locally:

INFERENCE_FAST_RELAY_URL=http://localhost:8888 \
INFERENCE_FAST_DASHBOARD_URL=http://localhost:3001 \
INFERENCE_FAST_API_URL=http://localhost:8788 \
bun run dev -- login

The dashboard and inference URLs are derived automatically from a localhost relay unless explicitly overridden.

Testing integrations

Three suites cover the harness integrations:

• bun run test:integration — hermetic; spawns the CLI against a local mock gateway and asserts the exact config every harness writes (all Claude slots, pickers, per-agent models, byte-exact off restore), plus drift between the shipped harnesses and the fast website's integration list. Runs in PR CI via ci:build-and-verify.
• bun run test:e2e (tests/auth-flow/) — the full auth flow against the LOCAL stack: a fresh account signs up, the CLI does the real device-code login (the test approves the code server-side, no browser), the relay mints a machine key through tRPC, and fast <harness> on writes config; logout restores and the disabled key is proven rejected by the gateway. Runs in the e2e harness ([inference] fast-cli auth-flow e2e job). Needs the local stack (task run-test-harness-headless) on current code — the CLI's catalog schema requires supported_endpoints on every model.
• .github/workflows/inference--fast-cli-e2e.yml — live end-to-end against prod (api.inference.net): installs the real harness binaries, connects them, and runs a two-turn probe per model — the second turn must reply and show a prompt-cache read covering most of the request (the "only the new message is uncached" invariant). Runs nightly against the latest @inference/fast@beta on npm. Prod because this is the release gate for the stack users actually hit, and many models are not hosted in dev. Its resolve step also carries the live-catalog contract (the CLI's default models must be present) so a serving-side catalog change fails the nightly run, not an unrelated PR.

To test any integration against any model(s) on demand:

gh workflow run inference--fast-cli-e2e.yml \
  -f harnesses=claude,codex \
  -f models=glm-5.2-fast,kimi-k3-fast \
  -f cli_source=npm-beta \
  -f enable_images=true

harnesses=all and models=default (the defaults) run every harness against its default models. A model that is not in the prod catalog fails the setup job before any install starts. The gateway converts every wire API (/v1/messages, /v1/responses, /v1/chat/completions) for every hosted model, so any catalog model can pair with any harness.

To run the same probe locally against prod with your own harness installs (seed a config.json with a machine key first, e.g. by running fast login against prod once and copying ~/.inference-fast/config.json):

INFERENCE_FAST_CONFIG_DIR=<state-dir-with-config.json> \
INFERENCE_FAST_RELAY_URL=https://relay.inference.net \
INFERENCE_FAST_API_URL=https://api.inference.net \
HARNESS=claude MODEL=claude-haiku-4-5 \
bun run test:live
# add ENABLE_IMAGES=true MODEL_VISION=true to include the image probe