Ficelle is an OpenAI-compatible router that puts the free models of 13 providers behind one endpoint. When one rate-limits or fails, the next eligible model answers in milliseconds. Paid fallback is off in code, so routed requests can't bill you.
One endpoint, 13 providers, free models only
Core includes OpenRouter + Nous; Pro maintains the broader pool Compare free tiers and limits →Free tiers rate-limit, retire models and change terms without notice. Ficelle treats that as the normal case: it measures, ranks and reroutes so your agent keeps getting answers.
A rate limit, an error or a failed capability check takes the model out of rotation. Ficelle tries the next eligible free model, cools the failed one down, and probes it until it recovers.
Why provider breakage kills agents →Ficelle checks whether each model really handles tool calls, JSON and vision, then ranks the working ones by measured speed and reliability. Only models that can run your request are eligible.
| model | tools | json | vision |
|---|---|---|---|
| deepseek-v4-flash | passes | passes | fails |
| nex-agi/Nex-N2-Pro | passes | passes | passes |
| glm-4-flash | fails | passes | fails |
Paid fallback is hardcoded off: flip it on and Ficelle refuses to start. A model that starts reporting a price is quarantined. The open core is public, so you don't have to take our word for it.
How strict-zero routing works →
Ficelle
allow_paid_fallback = False
# hardcoded, not a setting
config_store.py:96
The router runs locally. Your provider keys never leave it, and prompts go straight to the provider that serves them. There is no Ficelle proxy in the path.
http://127.0.0.1:8646/v1
Cursor, Codex, Continue, Open WebUI and your own scripts talk to Ficelle's OpenAI-compatible endpoint; Claude Code uses its native Anthropic endpoint. Hermes and OpenClaw use packaged connectors.
Call ficelle/auto-tools, auto-json or auto-fast. You say what the job needs; Ficelle picks the real model that delivers it.
Browse the model types →It scores working free models, routes to the best one available, and tries another eligible model when one fails. No manual provider switching.
Compare the providers →❯ ficelle demo --knock-out 2
Ficelle failover demo · profile ficelle/auto-tools
Prompt: "In one sentence, what is a rate limit?"
Candidates in the pool: 19
1. ficelle/nous/stepfun/step-3.7-flash:free (nous) — SIMULATED OUTAGE, HTTP 429 → rate_limited
(fabricated by the demo; no request was sent to this provider)
2. ficelle/nous/tencent/hy3:free (nous) — SIMULATED OUTAGE, HTTP 429 → rate_limited
(fabricated by the demo; no request was sent to this provider)
3. ficelle/openrouter/google/gemma-4-26b-a4b-it:free (openrouter) — answered, HTTP 200 · 1400 ms
Answered by ficelle/openrouter/google/gemma-4-26b-a4b-it:free in 1.4s.
"A rate limit is a restriction placed on the frequency of incoming requests to a resource, such as an API or a website, to prevent abuse and ensure system stability."
Cost of this request: $0.00
That is the whole product: the model you were about to use went down, and the request still came back.
Nothing was changed: no cooldown was applied, no route log was written.Change one base URL and keep the client-side API key as a placeholder. Your real provider keys stay inside Ficelle, on your machine. OpenAI-compatible clients need no adapter, Claude Code connects through the native Anthropic Messages endpoint, and Hermes uses a packaged connector.
from openai import OpenAI
client = OpenAI(
base_url="http://127.0.0.1:8646/v1",
api_key="ficelle-local",
)
client.chat.completions.create(
model="ficelle/auto-tools",
messages=[{"role": "user", "content": "..."}],
)curl http://127.0.0.1:8646/v1/chat/completions \
-H 'Content-Type: application/json' \
-d '{
"model": "ficelle/auto-json",
"messages": [{"role": "user", "content": "..."}]
}'export OPENAI_BASE_URL="http://127.0.0.1:8646/v1"
export OPENAI_API_KEY="ficelle-local"
# Tools that read these OpenAI environment variables now
# route through Ficelle without code changes.# Claude Code speaks Anthropic Messages; Ficelle serves it natively.
export ANTHROPIC_BASE_URL="http://127.0.0.1:8646"
export ANTHROPIC_API_KEY="ficelle-local"
claude # default mapping: Haiku -> auto-fast, Sonnet -> auto-coding, Opus -> auto-orchestratorIn Ficelle's 76-task benchmark, free routing was quality-safe against gpt-5 for labelling, extraction, reasoning, and code. Route those workloads through Ficelle and keep a paid model for polished writing.
Tagging tickets, detecting intent, sorting inbound, picking a branch in an agent graph.
ficelle/auto-fast
Free tied the premium model on every task we tested.
Pulling fields out of invoices, emails, scraped pages, logs. Anything that ends as JSON.
ficelle/auto-json
95% quality-safe in the benchmark.
Multi-step logic, small refactors, test scaffolding, working through a plan.
ficelle/auto-reasoning
The one we expected free to fail. It held up.
Launch copy, customer emails, release notes. Prose someone will judge you on.
keep your paid model
Paid genuinely writes better here. We are not going to pretend otherwise.
Verdicts come from a 76-task benchmark against gpt-5. See the numbers →
Long-form content is the honest exception: paid still writes better. The whole 76-task benchmark cost $0.68 to run.
The core router and its built-in profiles are free. Pro adds custom virtual models and maintains the broader provider pool alongside fusion, prompt compression, one-command updates and support.
Open core (BSL 1.1). No account, no card.
InstallBilled monthly. Cancel anytime. 109 € billed once a year.
Start 14-day trialOne payment. No renewals, ever.
Get the Founder license14-day free trial, card required. Cancel anytime before it ends. The 199 € Founder license is capped to the first 200. Routing stays local; Pro periodically reconnects to refresh its entitlement.
Clear answers on cost, privacy, model quality, and what happens when a provider goes down.
The open core is free under BSL 1.1. It includes strict-zero routing, OpenRouter and Nous reference providers, and the one-command installer. Pro adds the maintained provider pool, model fusion, prompt compression and one-command updates. You never pay for models through Ficelle, on any plan.
Ficelle marks the model as cooling down, tries the best working free alternative, and probes the limited one until it recovers. When another eligible model is available, your agent receives its response instead of the provider's 429.
For labelling, triage, structured extraction and most reasoning: yes. We benchmarked 76 real agent tasks against gpt-5 and free routing was quality-safe on classification (100%), JSON extraction (95%) and reasoning (94%). Polished long-form writing is the exception, where the paid model still wins. Route the first three through Ficelle and keep your budget for the writing.
No. There is no middleman: the router runs on your machine, your API keys stay local, and prompts go straight to the provider you choose. Ficelle-operated servers receive only an optional update request and the Pro license check; set FICELLE_DISABLE_UPDATE_CHECK=1 to opt out of update checks. The license check carries license identity, never prompts or provider keys.
No. Paid fallback is hardcoded off. It is not a setting you can mis-toggle, and Ficelle refuses to start if the config claims otherwise. Any model that reports billing gets quarantined automatically. The core code is public to inspect, so you can verify all of it.
Yes. The 14-day trial requires a card. Stripe charges the selected Pro plan when the trial ends unless you cancel first. Annual costs 109 €/yr; Lifetime costs 199 € as a single payment with no renewals. Prices are in euros, and your bank converts the EUR amount at its usual card rate.
Ficelle supports Cursor, Codex, Continue, Open WebUI, LangChain, CrewAI and plain curl through its OpenAI-compatible endpoint at http://127.0.0.1:8646/v1. Claude Code connects natively through Ficelle's Anthropic Messages endpoint: set ANTHROPIC_BASE_URL=http://127.0.0.1:8646. Hermes and OpenClaw use packaged connectors.
Ficelle Pro includes Profile Studio. You can create a stable ID such as ficelle/custom/coding, start from a built-in routing policy, tune its model pool and export it to Hermes, OpenClaw or a generic OpenAI-compatible client. If Pro expires, Ficelle preserves the configuration but stops listing, exporting and routing that custom ID; direct requests fail explicitly instead of silently falling back.
One command installs the free core, no account or card needed. Paid fallback is blocked by design.
curl -fsSL https://raw.githubusercontent.com/TheBlueHouse75/ficelle-open-core/v0.3.13/scripts/bootstrap-ficelle.py | python3 -
Ficelle Core v0.3.13, from its GitHub release. The script checks the package against a pinned SHA-256 before installing. Requires Python 3.
