model-preflight 0.1.9

Preflight checks for LLM prototypes.

ModelPreflight

Find out which cheap or free-ish LLM endpoints can carry your prototype before you wire them into your app.

ModelPreflight turns scattered provider keys into stable local groups like free_reasoning and free_fast, then lets you smoke-test prompts, fan out one-off questions, and fail over between providers without hard-coding model IDs everywhere.

If you want to...	Start here
See the free/dev endpoint menu	Free endpoint map
Try the payoff after setup	Run one check, then ask the pool
Get one green check quickly	60-second start
Try it without keys	No-key demo path
Run project smoke cases	Smoke tests
Use it as a Python helper	Library usage

---

ModelPreflight keeps provider setup machine-local and smoke cases project-local. It is not a hosted gateway, model leaderboard, or pricing oracle. It is the fast local preflight layer between "I found a promising free/dev endpoint" and "this provider is now wired into my product."

---

Free endpoint map

The high-value path is simple: collect provider keys once, let ModelPreflight group them, then ask free_reasoning or free_fast instead of memorizing every provider's model slug and quota page.

Provider	What it gives a prototype	Default group	Key env var	Setup
OpenRouter	Lowest-friction first run; one API key can route to free-tagged and paid models	free_reasoning	OPENROUTER_API_KEY	Auth docs
NVIDIA Build / NIM	High-capability open/open-weight hosted endpoints while the current dev access fits	free_reasoning	NVIDIA_NIM_API_KEY	API keys
Groq	Very fast repeated calls for fanout and smoke checks when free-plan limits fit	free_fast	GROQ_API_KEY	Console keys
Cerebras	Fast inference experiments for short prototype loops	free_fast	CEREBRAS_API_KEY	Inference docs
Mistral	First-party checks against Mistral model families	free_reasoning	MISTRAL_API_KEY	Account setup

The bundled presets are intentionally conservative starter data, not a claim that a provider will remain free, available, or quota-identical for every account. Provider catalogs, free tiers, and rate limits move; mpf doctor --live is the truth test for your machine today.

Secondary routes worth adding once the first pool works: Google Gemini/Gemma, Cloudflare Workers AI, GitHub Models, Hugging Face Inference Providers, and SambaNova. See docs/PROVIDER_PRESETS.md for the broader preset notes.

Run one check, then ask the pool

After one provider key is configured, first prove that the route works:

mpf demo

Shape of the output:

[
  {
    "id": "demo-ok",
    "passed": true,
    "failures": [],
    "text": "ok"
  }
]

Then ask a real one-off prompt with a single routed model call. Text output streams by default:

mpf ask "Write a poem about how ModelPreflight is the easiest way to try free LLM endpoints."

Shape of the output:

ModelPreflight finds the route,
checks the key, and sends it out...

Use pro when the prompt is worth asking several times. It fans out cheap samples, synthesizes the best answer through the reasoning group, and keeps an audit trail for which routes handled the calls.

mpf pro "Write the strongest short pitch for ModelPreflight Pro Mode: explain why fanout across cheap or free endpoints plus a judge pass is better than trusting one brittle LLM call for a prototype decision. Include one caveat." --n 8

Shape of the output:

{
  "final": "Pro Mode is useful when a prototype decision deserves more than one sample: fan out across cheap or free routes, compare independent answers, then synthesize the strongest result through a reasoning group...",
  "candidates": [
    {"index": 0, "ok": true, "text": "Fanout reduces single-sample luck..."},
    {"index": 1, "ok": true, "text": "The value is cheap parallel exploration..."}
  ],
  "group_winners": []
}

For structured-output work:

mpf pro "Design three robust JSON schemas for extracting vendor name, renewal date, total contract value, and termination notice from messy SaaS contracts. Include failure modes." --n 8

For repeatable project checks, write JSONL smoke cases once and run:

mpf init-project
mpf run

mpf demo proves the configured route works. mpf ask is for a single one-off prompt. mpf pro is for fanout plus synthesis. mpf run is for project-owned smoke files that should keep passing as prompts, providers, and model slugs drift.

For the snappiest CLI startup, install once with uv tool install model-preflight or pipx install model-preflight, then run mpf ... directly. uv run mpf ... may print package sync messages before ModelPreflight starts.

---

Why this repo exists

Early LLM prototypes often need a quick answer to a practical question: "Can this prompt, model group, or provider route work well enough to keep building?"

ModelPreflight gives you a lightweight layer for that stage:

• one global config for provider credentials and routing
• project-local JSONL smoke cases
• stable aliases such as free_reasoning and free_fast
• best-effort failover through LiteLLM
• audit records for live calls

When to use it

Use ModelPreflight when:

• a prototype needs cheap LLM smoke checks before deeper eval work
• several projects should share the same local provider setup
• you want logical groups instead of hard-coding provider/model IDs everywhere
• provider quotas, model slugs, or dev-tier availability may drift
• you need enough provenance to debug "which model answered this?"

What it is not

ModelPreflight is not:

• a model leaderboard
• a formal benchmark framework
• a hosted inference gateway
• a provider catalog authority
• proof that an endpoint is free, fast, or available today

Bundled provider presets are starter data. Check each provider's current catalog and terms before relying on a route.

60-second start

uvx model-preflight --help

# In a persistent tool or project environment:
uv tool install model-preflight
# or:
pipx install model-preflight

Set one supported provider key, initialize, and run one live check:

export OPENROUTER_API_KEY=...
# or: export NVIDIA_NIM_API_KEY=...
# or: export GROQ_API_KEY=...
# or: export CEREBRAS_API_KEY=...
# or: export MISTRAL_API_KEY=...
mpf init
mpf doctor --live
mpf demo

Expected signal:

• mpf init writes your machine-local config for the first visible provider key. If no supported key is visible, it writes the OpenRouter starter config and tells you to export OPENROUTER_API_KEY.
• mpf doctor --live prints a deployments table, then live check ok: group=....
• mpf demo prints JSON with "passed": true and an empty "failures": [] list.

Add checks to a project:

cd my-project
mpf init-project
mpf run

Expected signal:

• mpf init-project writes evals/smoke.jsonl, writes .model-preflight/README.md, and updates .gitignore.
• mpf run prints JSON results for the starter cases. Every passing case has "passed": true.
• A failing case exits non-zero and includes strings under "failures" so you know what drifted.

Both mpf and model-preflight are installed as console scripts.

ModelPreflight catches missing keys, broken provider routes, prompt formatting regressions, output-shape drift, accidental model/provider changes, and "this worked yesterday" prototype failures before you wire the LLM call into something larger.

No-key demo path

Use the minimal offline preset when you want to test the CLI and project workflow without a provider account:

mpf init --preset minimal
mpf doctor --live
mpf demo
mpf init-project
mpf run

What this proves:

• Config loading works without secrets.
• The CLI can run a live-style check through the offline echo provider.
• Project bootstrap works by creating evals/smoke.jsonl.
• Smoke scoring works when mpf run returns JSON where every case has "passed": true.

What it does not prove: remote provider auth, quota, latency, or model quality. Use the OpenRouter path below for that.

Install options

PyPI or isolated tool install

uv tool install model-preflight
# or:
pipx install model-preflight
mpf --help

Project dependency

uv add --dev model-preflight
# or:
pip install model-preflight

Editable checkout

git clone https://github.com/pylit-ai/model-preflight.git
cd model-preflight
uv pip install -e .
# or from another repo:
uv add --dev --editable /absolute/path/to/model-preflight

ModelPreflight requires Python 3.11+.

---

Machine-local config

ModelPreflight reads provider routes and secret-source references from your OS-specific user config directory by default. Use mpf paths to print the exact path. Override the path with either --config or MODEL_PREFLIGHT_CONFIG.

mpf paths
mpf init
mpf doctor
mpf models

With no --provider or --preset, mpf init checks visible environment variables in this order: OPENROUTER_API_KEY, NVIDIA_NIM_API_KEY, GROQ_API_KEY, CEREBRAS_API_KEY, MISTRAL_API_KEY. OpenRouter is only the fallback starter when none of those keys are visible. Explicit --provider and --preset always override auto-detection.

Provider keys are not stored in the config. For local cross-project use, link a machine-local dotenv file that stays outside this public package:

mpf setup --env-file /path/to/private/.env

Process env vars still win over linked dotenv values, which keeps CI and production behavior compatible with standard secret injection.

Provider setup is discoverable from the CLI:

mpf providers list
mpf providers guide nvidia
mpf providers guide openrouter
mpf providers test nvidia
mpf providers test openrouter

NVIDIA Build / NIM is the primary high-capability open/open-weight endpoint option. OpenRouter is still the lowest-friction discovery option because one API key can route to many model providers through an OpenAI-compatible API.

Use either primary path:

mpf setup --env-file /path/to/private/.env
mpf doctor --group free_reasoning --live

mpf init --provider openrouter
export OPENROUTER_API_KEY=...
mpf doctor --provider openrouter --live

For agent and CI readiness checks, make sure provider keys are visible in the agent process environment or through a linked machine-local secret source, then use JSON diagnostics:

mpf doctor --group free_reasoning --json

status: "ok" means config and required keys are present. error_code distinguishes MISSING_REQUIRED_ENV, GROUP_NOT_FOUND, and disabled matching provider/group cases.

Provider	Best for	Env var	Setup
NVIDIA Build / NIM	Primary high-capability open/open-weight endpoint pool	NVIDIA_NIM_API_KEY	API keys
OpenRouter	One-key first run with broad model access	OPENROUTER_API_KEY	Authentication docs
Groq	Fast repeated calls after first-run setup works	GROQ_API_KEY	Groq console
Cerebras	Fast inference experiments when current dev-tier limits fit	CEREBRAS_API_KEY	Cerebras inference docs
Mistral	First-party Mistral model-family smoke checks	MISTRAL_API_KEY	Mistral API keys

Secondary/overflow pool to add manually once the primary pool works: Google Gemini/Gemma, Cloudflare Workers AI, GitHub Models, Hugging Face Inference Providers, and SambaNova. These are documented in docs/PROVIDER_PRESETS.md, but not packaged as first-run presets yet because auth shape, model IDs, and free/dev limits are more account-specific.

The default config creates logical groups, then maps each group to one or more LiteLLM deployments:

router:
  num_retries: 1
  timeout_seconds: 60
  default_group: free_reasoning
  audit_jsonl: null
artifacts_dir: ~/.cache/model-preflight/artifacts

deployments:
  - name: nvidia_nim_nemotron_3_super
    provider: nvidia
    group: free_reasoning
    model: nvidia_nim/nvidia/nemotron-3-super-120b-a12b
    api_key_env: NVIDIA_NIM_API_KEY
    enabled: true
    required: true
    status: best_effort
    setup_url: https://build.nvidia.com/settings/api-keys
    rpm: 10
    tier: reasoning

Provider preset discipline

Provider presets are best-effort starter data, not authoritative claims about free availability.

• user-local config wins over bundled defaults
• mpf doctor fails fast when required keys are missing
• optional/disabled providers do not block first-run checks
• live checks should be opt-in in CI
• endpoint names, quotas, pricing, and behavior can change without this repo knowing

See docs/PROVIDER_PRESETS.md for the preset rules.

Custom config path

mpf init --config ./model-preflight.yaml
mpf doctor --config ./model-preflight.yaml
mpf doctor --config ./model-preflight.yaml --live

export MODEL_PREFLIGHT_CONFIG="$PWD/model-preflight.yaml"
mpf models

Use environment variables for secrets. Do not commit provider keys.

If you use 1Password, see docs/secrets/1password.md for linked dotenv and op run examples. Run mpf init --provider <provider> once to create the machine-local provider config.

---

Smoke tests

Smoke cases are JSONL files owned by the project that is doing the prototype work.

{"id":"basic-ok","prompt":"Return only: ok","expected_substrings":["ok"]}
{"id":"avoid-word","prompt":"Answer yes without using the word nope","forbidden_substrings":["nope"]}

Run them with:

mpf run
# or:
mpf run path/to/smoke_cases.jsonl

mpf run prints JSON results and exits non-zero if any case fails.

Case fields

Each smoke case supports:

• id: stable case identifier
• prompt: user prompt sent to the configured model group
• group: optional model group override
• expected_substrings: strings that must appear in the answer
• forbidden_substrings: strings that must not appear in the answer

These checks are intentionally simple. They are meant to catch obvious routing, prompt, and regression problems before you spend time on heavier evals.

---

Ask

mpf ask sends one prompt through one configured model group and prints the model text to stdout. Plain text streams as tokens arrive. Progress and route metadata go to stderr by default, so stdout stays clean for pipes and command substitution. In an interactive terminal, stderr status lines are styled and separated from the answer by a blank line. Use --quiet to suppress all stderr status lines, or --hide-route to hide provider/model route metadata while keeping progress visible. JSON output is buffered so it remains valid JSON and includes route metadata unless --hide-route is set.

mpf ask "Write a poem about how ModelPreflight is the easiest way to use free LLM endpoints."
mpf ask "Write a shell-safe tagline" --quiet
mpf ask "Which model route is this using?"
mpf ask "Keep route metadata hidden, but show progress" --hide-route
mpf ask "Summarize why free endpoint preflight matters" --no-stream
mpf ask "Return JSON only: {\"ok\": true}" --group free_reasoning --json

Use ask for quick manual checks, demos, and shell snippets. Use run when the same prompt should become a repeatable smoke case.

---

Pro Mode

mpf pro fans out a one-off prompt, then synthesizes a final answer through a judge group.

mpf pro "Use fanout plus synthesis to choose a robust JSON schema strategy for extracting renewal terms from messy SaaS contracts. Return the final schema, validation rules, and the main failure mode." --n 8

Defaults:

Option	Default	Role
--n	8	number of sampled answers
--sample-group	free_fast	fanout group
--judge-group	free_reasoning	synthesis group

Cost and quota note

Fanout multiplies live provider calls. Keep --n low while testing, use restricted provider keys where available, and review provider dashboards when running against paid endpoints.

ModelPreflight records audit rows for live calls, but it does not enforce provider billing limits beyond your configured routing and provider-side controls.

---

Library usage

from model_preflight import ModelGateway, load_config, pro_mode

gateway = ModelGateway(load_config())

print(gateway.text("Return only: ok", group="free_reasoning"))

result = pro_mode(gateway, "Solve this toy puzzle", n=8)
print(result["final"])

The library API is intentionally thin:

• load_config() reads the same machine-local config as the CLI
• ModelGateway wraps LiteLLM Router with stable group aliases and audit logging
• pro_mode() runs fanout plus synthesis for one-off prototype prompts

---

Audit artifacts

By default, ModelPreflight writes audit logs under:

~/.cache/model-preflight/artifacts/audit.jsonl

Each live call should be traceable enough to debug provider drift:

• timestamp
• logical group
• resolved provider/model when returned by the provider
• prompt or case metadata
• latency
• token usage when available
• response id when available

See docs/EVAL_PROVENANCE.md for provenance expectations.

---

Repo adapters

Path	Purpose
examples/autoharness_provider.py	Drop-in provider wrapper for AutoHarness-style experiments
examples/gpt_pro_mode_refactor.py	Example refactor from single-provider Pro Mode to shared routing
examples/node_hook_example.mjs	CLI bridge for JS or agent-hook projects
skills/model-preflight/SKILL.md	Optional coding-agent skill for consistent usage

Command reference

mpf init --provider openrouter
mpf doctor --live
mpf demo
mpf ask "write a tiny launch blurb for ModelPreflight"
mpf init-project
mpf run
mpf providers list
mpf providers guide openrouter
mpf models
mpf pro "solve this toy task" --n 8

Contributor workflow

uv sync
uv run pytest
uv run ruff check .
uv run mypy src

Package metadata lives in pyproject.toml. Tests live under tests/.

---

Design principles

• Global provider routing lives in the path printed by mpf paths.
• Project-local checks define cases, scoring, fixtures, and artifacts.
• LiteLLM handles provider-specific API quirks.
• ModelPreflight adds stable aliases, lightweight failover, and audit logs.
• Deterministic tests should run before live provider checks.

For the product scope and non-goals, see docs/NORTHSTAR.md.