freellmpool

Pool the free tiers of 19 LLM providers cataloged in freellmpool (237 enabled chat routes, 358 cataloged chat models) behind one OpenAI-compatible endpoint — as a CLI, a Python library, or a local proxy. Can start without API keys when a keyless provider is up.

FAQ: where prompts go, ToS posture, failover, bans, and comparisons.

30-second quickstart

Fresh install to first free-model reply is measured at about 19 seconds under the 30-second target on a clean Linux/Python 3.12 environment, with no API keys when a keyless provider is up:

python3 -m venv .venv
. .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install freellmpool
freellmpool ask --max-tokens 32 "Reply with one short sentence: freellmpool is ready."

CI runs the same path from this checkout with FREELLMPOOL_QUICKSTART_PACKAGE=. scripts/quickstart-test.sh.

Groq, Cerebras, NVIDIA NIM, Google Gemini, OpenRouter, GitHub Models, Cloudflare, Mistral, Cohere and others each give away a free tier — but each has its own SDK, rate limits, and daily cap. freellmpool puts them in one pool: it sends each request to a provider you have access to, fails over to the next when one is rate limited or down, and tracks per-day usage so you get the most out of every tier.

Several providers (Pollinations, OVHcloud, and Kilo Gateway) need no API key, and LLM7 works without one, so the quickstart can answer without signup when a keyless provider is available.

To inspect your local provider keys, agent CLIs, proxy config, and Tailscale state before wiring tools, run the print-only init wizard:

freellmpool init --yes
freellmpool init --yes --agent opencode
freellmpool init --yes --agent metaswarm --tailnet

Add keys for the other providers to unlock more models and higher limits.

First-run setup with freellmpool init

freellmpool init inspects provider keys, installed agent CLIs, Tailscale state, and proxy config, then prints one copy-pastable next step without editing files. Run it detect-only first:

freellmpool init --yes

--json emits the same detection as versioned JSON for scripts and agents.

Tailnet / remote agent gateway

Serve the proxy on your Tailscale 100.x address with a generated API key:

freellmpool tailnet serve --port 8080

From a remote machine:

freellmpool tailnet connect <tailnet-ip> --port 8080

Both sides support --api-key <shared-secret> if you want to pin a key instead of using a generated token. Tailnet serving requires auth by default; do not run unauthenticated over non-loopback interfaces.

Metaswarm agent lanes

This project uses one Umans/Kimi K2.7 worker lane, one MiniMax M3 lane, Codex as escalation, and Claude Opus only for final pre-ship review. The installable Metaswarm profile mirrors that posture: one free/cheap worker lane through the local proxy, one larger freellmpool reviewer lane, and Codex/Opus as explicit user-owned paid escalation/final-review lanes only (never silent).

freellmpool init --yes --agent metaswarm --tailnet
freellmpool profile install metaswarm
freellmpool tailnet serve --port 8080
freellmpool profile doctor metaswarm --dry-run

Run a coding agent on free models

freellmpool's proxy speaks the OpenAI API and includes an experimental Anthropic-compatible path, so coding agents can run against pooled free tiers — just point them at the proxy:

freellmpool proxy                       # starts http://localhost:8080
freellmpool code claude                 # prints the one-line setup for Claude Code
freellmpool profile list                # richer installable profiles
freellmpool profile show metaswarm      # Tailnet-aware Metaswarm profile
# (also: codex, aider, cline, continue, cursor, opencode, metaswarm)

Claude Code gateway mode can also be launched directly:

ANTHROPIC_BASE_URL=http://localhost:8080 \
ANTHROPIC_AUTH_TOKEN=dummy \
ANTHROPIC_API_KEY=dummy \
ANTHROPIC_MODEL=auto \
ANTHROPIC_SMALL_FAST_MODEL=auto \
CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1 \
claude

Existing OpenAI-compatible apps work the same way: set OPENAI_BASE_URL=http://localhost:8080/v1 and keep your code unchanged. Anthropic-compatible tools can use the experimental bridge with ANTHROPIC_BASE_URL=http://localhost:8080.

OpenCode gets a deeper integration: a live in-editor dashboard (routing mode, estimated savings, tokens served free, provider race, latency), per-request quality routing via the model picker (freellmpool/auto|fast|quality|fair), and freellmpool_status / freellmpool_models tools — see integrations/opencode-tui and the guide.

New in 0.11: capacity tools — freellmpool capacity status shows which free tiers are usable right now, freellmpool providers health live-probes them, and freellmpool keys add walks you through configuring more (see Capacity & provider health and docs/CAPACITY.md).

New in 0.10: an async API (AsyncPool), an MCP server (freellmpool mcp), latency-aware routing with freellmpool benchmark, observability hooks, and a plugin system for custom providers. See the changelog.

Install

pip install freellmpool      # or: pipx install freellmpool

Only dependency is httpx. Python 3.11+.

Command line

freellmpool ask "Write a haiku about sqlite"
git diff | freellmpool ask "Write a commit message for this"
freellmpool tokenmax "Hardest question you've got"  # 🌈 blast models, print answers, optional synthesis
freellmpool providers        # which providers are configured
freellmpool models           # every provider/model id
freellmpool stats            # lifetime tokens served free + estimated cost avoided
freellmpool badge -o badge.svg   # a shareable SVG badge of that total

freellmpool tokenmax is the tongue-in-cheek maximum-effort mode: it fans your prompt out to many available models at once and prints each answer. The CLI adds a synthesized verdict by default unless you pass --no-synthesize; the MCP tool returns the model answers for the calling agent to synthesize. (See docs/MCP.md.)

freellmpool stats is a running, persistent lifetime total (it survives restarts and upgrades). Embed freellmpool badge in a README, or serve it live from the proxy at /badge.svg (set FREELLMPOOL_PUBLIC_BADGE=1 to make it publicly embeddable).

Pin a provider or model; common OpenAI/Anthropic model names are mapped to a free equivalent so existing scripts keep working:

freellmpool ask -m groq/llama-3.3-70b-versatile "hi"
freellmpool ask -p cerebras,groq "hi"
freellmpool ask -m gpt-4o-mini "hi"      # routed to a free model

Roles

freellmpool roles lists ask-role presets (coder, critic, summarizer, long-context, cheap, fast, second-opinion, ...). Each role sets routing, token budget, temperature, and system-prompt hints without inventing a second routing engine. Explicit flags (--model, --providers, --routing, --max-tokens) win over role defaults, and the verbose output shows when an override happened.

freellmpool ask --role coder "write a pytest for this function"
FREELLMPOOL_MODE=wise freellmpool ask --role cheap "summarize this patch"

As a proxy

Run a local server that speaks the OpenAI API, then point any OpenAI-compatible tool at it. On loopback, any placeholder API key works unless you configured FREELLMPOOL_PROXY_KEY or passed --api-key; Tailnet/LAN serving requires a real proxy bearer token by default.

freellmpool proxy
export OPENAI_BASE_URL=http://localhost:8080/v1
export OPENAI_API_KEY=unused

from openai import OpenAI
client = OpenAI()
print(client.chat.completions.create(
    model="auto",
    messages=[{"role": "user", "content": "hi"}],
).choices[0].message.content)

# audio → text (Whisper), same client:
print(client.audio.transcriptions.create(
    model="auto", file=open("audio.mp3", "rb"),
).text)

Or with curl (multipart upload):

curl -s http://localhost:8080/v1/audio/transcriptions \
  -F file=@audio.mp3 -F model=auto

The proxy also implements the OpenAI Responses API (for the Codex CLI) and an experimental Anthropic Messages API path (for Claude Code), so coding agents can run on free models too. freellmpool code <agent> prints the exact setup, while freellmpool profile install <agent> prints the fuller copy-pastable profile without mutating third-party config:

freellmpool code aider       # also: claude, codex, cline, continue, cursor, opencode
freellmpool profile show opencode
freellmpool profile doctor opencode --dry-run

Main proxy surfaces:

• /v1/chat/completions — OpenAI-compatible chat, token streaming, tool calling.
• /v1/responses — minimal Responses API shim for Codex-style agents.
• /v1/messages — experimental Anthropic-compatible Messages path.
• /v1/embeddings and /v1/audio/transcriptions — OpenAI-compatible embedding and Whisper-style multipart transcription.
• /v1/models — routing aliases plus concrete provider/model ids.
• /freellmpool/battle and /playground — bounded browser/JSON model comparisons.
• /dashboard, /status, /healthz, /badge.svg — local operations surfaces.

/playground and the API routes are auth-protected when the proxy key is set. Setup snippets for specific tools are in docs/INTEGRATIONS.md and docs/AGENTS.md. The repo also includes an experimental metaswarm review adapter for using freellmpool as an external-tools reviewer/second opinion. freellmpool profile show metaswarm documents a free/cheap worker lane, a larger reviewer lane, Tailnet client setup, and paid Codex/Opus lanes as explicit user-owned escalation paths only.

As a library

from freellmpool import Pool

pool = Pool.from_default_config()
reply = pool.ask("Summarize the plot of Hamlet in 20 words.")
print(reply.text, "—", reply.provider_id)

vectors = pool.embed(["first document", "second document"]).vectors

with open("audio.mp3", "rb") as f:
    text = pool.transcribe(f.read(), "audio.mp3").text   # Whisper, failover across providers

Async is the same API with await:

from freellmpool import AsyncPool

async with AsyncPool.from_default_config() as pool:
    reply = await pool.aask("Summarize the plot of Hamlet in 20 words.")

Pass on_event=... to either pool to receive structured routing/cache events (attempt/success/error/cooldown/cache_hit/cache_miss/exhausted) for logging or tracing. Add your own endpoint with register_provider(...), or a new request shape with register_adapter(name, fn).

Benchmark your providers

freellmpool benchmark times one call per configured provider and prints latency and success, so you can see which of your free tiers are fastest right now. The router learns the same latency/success signal from real traffic as it runs; set FREELLMPOOL_ROUTING=fast to prefer the lowest-latency provider instead of the default least-used-first.

$ freellmpool benchmark
  provider/model            status   latency  note
  cerebras/llama-3.3-70b    ok        180 ms  6 tok
  groq/llama-3.3-70b        ok        240 ms  6 tok
  ovh/Meta-Llama-3_3-70B    FAIL           -  HTTP 429

Capacity & provider health

Free tiers drift through the day — keys expire, providers go down, daily caps fill. These commands tell you what's usable right now and what to set up next:

freellmpool capacity status --target 5   # who's healthy / near quota / missing a key
freellmpool quota-wise status            # local headroom + recommended mode
freellmpool providers health             # send one tiny request to each, time it
freellmpool keys checklist --target 5    # which keys to add to reach N healthy providers
freellmpool keys add groq                # configure a key (and record metadata)

capacity status is local-first: it reads your catalog, environment, and per-day quota counters and labels each provider healthy, low_quota, exhausted, invalid_key, or missing. It also syncs an advisory external catalog (mnfst/awesome-free-llm-apis) to suggest free providers you could add — advisory only; your providers.toml stays the source of truth for routing. keys add <name> can even import a suggested provider from that catalog or create an OpenAI-compatible stub and autodiscover its models. The proxy /dashboard shows the same capacity at a glance. Full reference: docs/CAPACITY.md.

FREELLMPOOL_MODE=wise is the conservative quota mode: ask defaults to a smaller output budget and spread routing, tokenmax narrows its default fan-out, and broad multi-model calls require confirmation unless you pass --yes. Per-command --mode normal|wise overrides the environment, and [settings] mode = "wise" works from config.toml. The conserve role is a quota-conscious shorthand for small, spread-routed answers.

For a bounded second opinion instead of a full tokenmax blast:

freellmpool ask --second-opinion --opinions 3 "is this implementation plan sound?"
freellmpool ask --role second-opinion --synthesize "which release note is clearer?"

The shared panel asks a few diverse providers, keeps individual failures visible, and can append a non-fatal synthesis when you pass --synthesize.

For a side-by-side comparison you can inspect in the terminal or local browser:

freellmpool battle "which changelog entry is clearer?" --synthesize
freellmpool proxy --port 8080
freellmpool playground --port 8080

Bundled recipes wrap common workflows in JSON files you can inspect and run:

freellmpool recipe list
freellmpool recipe run second-opinion "is this launch plan clear?" --synthesize
freellmpool recipe run pr-review --input patch.diff
freellmpool recipe run repo-summary --path 'src/freellmpool/*.py'
freellmpool recipe run metaswarm-worker-review --input worker.md --validation-output-file validation.txt

Recipes use the same role presets and shared panel helper as ask and battle; there is no separate routing engine.

Local foreground job queue

For slow, quota-aware work that should not block a live session, queue jobs to an append-only JSONL log under your config dir (override with FREELLMPOOL_JOBS_PATH). The queue is foreground-only: jobs run processes one job at a time and records started/completed/failed/cancelled events. Completed ask jobs keep their output in the job log; completed recipe jobs also write run records and Markdown reports via the same report helpers used by freellmpool report.

# queue a recipe job
freellmpool jobs add --recipe pr-review --input patch.diff

# queue an ask job with a role preset
freellmpool jobs add --role summarizer "summarize the latest changelog"

freellmpool jobs list            # replayed state (idempotent across restarts)
freellmpool jobs watch           # one-shot refresh render, no daemon

freellmpool jobs run --dry-run   # print execution order, mutate nothing
freellmpool jobs run --max-failures 2   # halt after N consecutive failures
freellmpool jobs cancel <job-id> # append a cancel tombstone, not a mutation

freellmpool report list
freellmpool report last --markdown
freellmpool report last --html --path
freellmpool cost show <run-id>

Cancellation is a new tombstone event, not a re-write of the earlier queued record — a crash before jobs run finishes still leaves the queue replayable, and cancelled jobs stay cancelled after restart. Duplicate submissions create distinct jobs; pass --dedupe to reject re-submission of the same recipe or role while a job is still pending.

As an MCP server

freellmpool mcp runs a Model Context Protocol server over newline-delimited JSON-RPC stdio, so Claude Desktop, Claude Code, or Cursor can hand subtasks to free models. It exposes ask, panel/second-opinion, battle, recipe, roles, Tailnet-info, quota-wise, route-preview, models, quota, stats, and tokenmax tools. See docs/MCP.md. A server.json is included for the MCP registry.

In Simon Willison's llm CLI

There's a plugin: llm install llm-freellmpool → llm -m freellmpool "..." with no API key. Source: 0xzr/llm-freellmpool.

Provider keys

freellmpool reads keys from the environment and uses whatever is set. None are required. Step-by-step signup links for each (all free, no card) are in docs/ACCOUNTS.md.

A config.toml (see config.toml.example) can hold keys, model aliases, and settings instead of env vars.

Local diagnostics and operations

Run freellmpool doctor for a no-network local check of package version, config paths, configured provider count, routing mode, quota/cache locations, external catalog cache age, and bundled catalog validity.

Response caching is off unless FREELLMPOOL_CACHE_TTL (seconds) or [settings] cache_ttl is positive. When enabled, cache rows live in SQLite with WAL mode and TTL pruning; FREELLMPOOL_CACHE_MAX_ENTRIES caps retained rows (default 10000, set 0 to disable size pruning).

Quota counters are written immediately by default. Long-running proxy/MCP processes can reduce file churn with FREELLMPOOL_QUOTA_FLUSH_EVERY=N, which batches up to N successful requests before flushing. Shutdown paths and quota.snapshot() flush pending counts, so dashboards and process exits still see current totals.

How routing works

For each request, freellmpool builds the list of (provider, model) pairs you have access to, then orders providers least-used-first and picks a least-used model inside that provider. This keeps providers with large catalogs, like NVIDIA, from receiving more traffic only because they expose more models. A provider that returns a 429 is set aside for a cooldown window. Daily counts are kept in ~/.config/freellmpool/quota.json and reset at UTC midnight.

Every call records latency and success per model target. A provider whose targets are currently failing sinks to the back automatically; with FREELLMPOOL_ROUTING=fast the fastest measured provider goes first instead. FREELLMPOOL_ROUTING=fair spreads requests across providers to preserve daily quota. freellmpool benchmark warms these metrics on demand. To restore the old per-model balancing behavior, set FREELLMPOOL_ROUTING=legacy or FREELLMPOOL_ROUTING=model (or FREELLMPOOL_ROUTING=model-fast for the old per-model fastest-first ordering).

Quality routing (FREELLMPOOL_ROUTING=quality). Free tiers' strongest models have the smallest daily caps, so a naive pool gets weaker as the day fills. Quality routing matches each prompt's difficulty to each model's capability: hard prompts (long input, code, reasoning cues) go to the strongest available model, and easy ones go to lightweight models — which rations scarce strong-model quota so the pool stays sharp for longer. Capability is grounded in real benchmark data, not guessed from names; models that no benchmark lists cover fall back to a name heuristic.

The bundled, offline scores come from LMArena Elo (an MIT-licensed snapshot) and the Aider code-editing leaderboard (Apache-2.0), normalized to a common percentile scale. For much broader coverage, run freellmpool capability sync with a free Artificial Analysis API key (FREELLMPOOL_AA_API_KEY) — its Intelligence Index covers most current and open-weight models and takes precedence. The fetched AA data is cached locally under your own key (never bundled, per AA's terms). freellmpool capability status shows current coverage. Scores via LMArena and Aider; intelligence index via Artificial Analysis when keyed.

Context windows. Free models often have small context windows. freellmpool never truncates your input; instead, when a model rejects a request as too long, it learns that model's limit and stops routing oversized requests there, escalating only to larger-window models. If nothing fits it raises a clear ContextWindowExceeded (with the estimated input size) instead of a generic failure — over the proxy that's a 413. You can declare a model's window with context = N in providers.toml to skip it proactively.

Architecture notes: docs/ARCHITECTURE.md.

Limitations

• Free-tier models are smaller than frontier models. They're good for drafting, summarizing, classification, triage, and everyday coding — not a replacement for GPT-class reasoning on hard problems.
• Quality and capacity vary through the day as high-cap tiers exhaust; limits reset at UTC midnight.
• Free tiers change without notice. When a model id or limit goes stale, a one-line PR to providers.toml fixes it for everyone.
• The proxy is meant for local/single-user use. It binds to 127.0.0.1 by default; if you expose it, set a key (--api-key).
• The Claude Code / Anthropic path is experimental (text and tool use; no vision).
• These are free tiers shared by everyone — don't abuse them.
• Upstream providers receive and may prompt-title/moderate your requests; use the FAQ privacy table and each provider's current terms before sending sensitive prompts.
• Free-tier availability, model IDs, and rate limits drift without notice. freellmpool does not bypass provider limits, rotate accounts, or evade quotas.

How it compares

FreeLLMAPI predates this project, and the overlap is independent convergence: both projects noticed that legitimate free tiers are useful when treated carefully. freellmpool's niche is the keyless, pip-installable client for squeezing hosted free tiers from a CLI, library, local proxy, and MCP server; OpenRouter is the polished hosted route; LiteLLM is the mature bring-your-own-key gateway.

Table sources: freellmpool's catalog and proxy code in this repo; OpenRouter's quickstart, free-model, routing, and audio docs; LiteLLM's README, MCP docs, and audio transcription docs; FreeLLMAPI's README.

FAQ

Is there a free, OpenAI-compatible LLM API gateway? Yes — freellmpool is a free, MIT-licensed gateway that exposes one OpenAI-compatible endpoint backed by the free tiers of 19 cataloged providers. pip install freellmpool and point any OpenAI client at the local proxy.

How do I use multiple free LLM APIs at once? freellmpool pools them: each request goes to a provider you have access to, fails over to the next when one is rate-limited or down, and tracks per-day usage so load spreads across tiers.

Can I run Claude Code or Codex on free models? Yes — the proxy speaks the OpenAI API and has an experimental Anthropic-compatible path. Set OPENAI_BASE_URL=http://localhost:8080/v1 for OpenAI-compatible tools or ANTHROPIC_BASE_URL=http://localhost:8080 for Anthropic-compatible tools, then run Codex, Claude Code, aider, Cline, Continue, or Cursor against pooled free tiers. For Claude Code, set CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1 so /v1/models is discovered through the Anthropic bridge. See freellmpool code <agent>. (Claude Code path is experimental: text + tools, no vision.)

Do I need an API key? No — Pollinations, OVHcloud, and Kilo Gateway work with no key, and LLM7 is key-optional, so a fresh install can answer without signup when a keyless provider is available. Add free keys for the other providers for more models and higher limits.

Is it free and open source? Yes, MIT-licensed. More at the project page.

Featured in

• Community videos (Spanish, by lytohlg AI): "Accede a 18 modelos de IA GRATIS con 1 solo comando" and "Prueba 18 IAs GRATIS sin API key en 30 segundos" (from an earlier catalog; freellmpool now catalogs 19 providers).
• Directory: FreeLLM Pool on MCP Market.

Contributing

New providers and fixes to stale limits are the most useful contributions, and both are usually a small change to providers.toml. See CONTRIBUTING.md. Maintainer-ready newcomer tasks are drafted in docs/GOOD_FIRST_ISSUES.md. Tests run with no network access:

python -m pip install -e ".[dev]" && ruff check . && pytest

Source-first verification in this repo uses PYTHONPATH=src so pytest exercises the checkout without requiring an editable install first; CI runs the same configuration. Release readiness uses PYTHONPATH=src python3 scripts/check_release_ready.py.

License

MIT