Pick a tool, skill or branch from a list
Four or more companies' docs show a full fit.
What counts as done
One option from the caller's list, with a 'none' option available when no option fits.
Measured, for our own tests: the option a human would pick, with 'none' chosen when no option fits, on at least 45 of 52 held-out routing cases, under 300 ms hosted. No docs we read state these figures, so they don't count against a listing here.
How agents ask for it: which tool should I call for this
, which skill applies
, do I have enough information to act
, which document entry matches this form field
Does the job 12
The docs show it meets the test above, and nothing we read says otherwise.
| Listing | Why |
|---|---|
| Celeris-1 Decision D Marqo Inc | choice with up to 512 options and a probability per option, so a none option can be added; confidence is the rescaled top probability. No documented none-of-these pattern and no latency figure. source |
| Clef B Cloudflare, Inc. | Arize's 52-case tool and skill routing, including none-of-them cases: Clef 27B got 50 of 52, the best score. 64 questions a call. source |
| Decision 3.0 B vLLM Semantic Router project | choice over named options with permutation_average available in the runtime to remove order dependence; many questions in one call. No independent routing set was found. source |
| elephant-1 D Baseten Labs, Inc. | choice over up to 255 options with a trained 'I don't know' option, all questions sharing one prefill; the report lists tools and automation among its five areas and the card beats Jev 1.14 there. No independent routing set. source |
| Kev BB Jared Palmer | Arize's 52-case routing final with none-of-them cases: Kev 9B 48 of 52, tied with Jev. /v1/systemone/permute runs a choice with up to 64 option orders. source |
| Liquid d1 C Liquid AI, Inc. | Arize's 52-case routing final: Liquid d1 got 48 of 52. choice with probabilities and a routing example in the guide. source |
| mammoth-1 D Baseten | Top of the Decision Index 0.4 at 71.24 and leads four of five areas including tools and automation; the community benchmark's 84.9 percent over 31 categories at 240 ms median, weak on exact counting and procedures. source |
| Drex 1.5 BB Nace AI, Inc. | choice with up to 255 options on Max and an 'other: null' option in the Questions guide's example; confidence is the average of the model's scores. No order control. source |
| OpenAI Decisions API A OpenAI OpCo, LLC | Arize's 52-case routing final: 48 of 52, with the 'other' fallback the guide documents for options that cover nothing; images can sit in question instructions. source |
| pplx-decider BB Perplexity AI, Inc. | One choice question with up to 255 options and an 'other' option; the docs' tool-gate and browser-agent cookbooks use this shape. Send candidate tools as one choice, not separate questions. source |
| kettei C Sqwish Ltd, trading as Sqwish Labs | single decisions with 2 to 255 outcomes, 'none' included, and an abstain action when p_top is under your threshold; the docs' tool-call example routes to run, ask or refuse. source |
| Jev B TypeSafe AI, Inc. | Arize's 52-case routing final: 49 of 52 with none-of-them cases; the companion-noul pattern detects when nothing fits; a skill-suggestion cookbook and a user's branch choices inside a skill. source |
Does the job, with a limit 13
It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.
| Listing | Why |
|---|---|
| Decider BB Mark Marosi (Mapika) | Issue 9's 167-item set of tool routing, navigation and form filling scored 132 of 167 on the 2B, which clicked 'Start over' at 0.66 confidence. Larger models score higher on the Decision Index tools area but weren't tested on that set. source |
| Drex Max v1 C Nace AI, Inc. | choice with 255 options and an 'other: Anything else' option in the card's example request. No order control; licence cap. source |
| GLiClass C Knowledgator Engineering Ltd. | Label selection over up to 128 names in one pass, but no documented none-of-the-above route; issue 29 asks how to train one and is unanswered. source |
| Granite Switch 4.1 B International Business Machines Corporation | Options are single-token labels, 128 per request before chunking; no typed none option or order control. Caller computes the softmax. source |
| JEV-27B-VL C AutoTrust AI Lab | Up to 256 options, but beyond 16 options the head uses labels it never saw ('wide-labels' adaptation) and one request holds one question. source |
| joey-1 C Baseten Labs, Inc. | choice up to 255 options with a trained 'I don't know' option, but the 2B scores 10.2 on the index's new-domain tests, so tool lists unlike its training are a risk. source |
| Microsoft-Decision-1 E Microsoft | choice with probabilities, with 'use a low top probability to escalate' guidance and skill-based decisions named as a use; OpenRouter's 0.42 s is over the 300 ms target. source |
| pplx-decider-flash B Perplexity AI, Inc. | choice up to 255 options with the 'other' pattern, but the 4B trails the 27B and the docs tell you to compare both on labelled data at your thresholds first. source |
| RightNow System Zero B RightNow AI, Inc. | choice options are bounded by limits.choice_options from the model list, and the docs say text, PDF and video allow 9 tools and none (image or audio 25), which is under the 2 to 50 the use case needs for larger tool lists. source |
| ScaleDown E Carbon ScaleDown, Inc. | choice over criteria options, but the classify path's 20-label cap and the undocumented options cap on /v1/scaledown leave larger tool lists unverified; no none pattern documented. source |
| SPEV D Sprag AI, Inc. | choice up to 255 options with a refusal type but no documented none option or order control; latency unpublished. source |
| Strands Decider 2B B Amazon Web Services, Inc. | Arize routing 45 of 52, precisely the bar; AWS Builder's tool-selection desk and routing AUC 0.51 against model routing 72 of 80 elsewhere. First-option preference measured at 55 percent. source |
| Vela 2.0 BB vLLM Semantic Router project (with KR Labs) | choice and set questions over named options; the 4B's general decision index of 31.63 and no order control make a 45-of-52 bar doubtful without measurement. source |
Doesn't do this job 1
Listings in the category that do something else, with the reason.
1 listings in the category were added after this reading and aren't marked yet.
What stops agents finishing
- first-option bias on small models flips the pick on order swap
- tool descriptions plus state exceed an 8k or 16k window
- confidence rarely reaches 0.9, so a strict threshold never fires
Around it
Before: agent-frameworks.plan-step, mcp-servers.list-tools. After: mcp-servers.call-tool, agent-frameworks.load-skill.
Where this comes from
Weight 10 of 100 in Decision models: 5 of 41 accounts; Arize's 52-case routing final, the mager.co prototype, the decider 167-item set and the AWS tool-selection desk
The same as data: /api/v1/use-cases/decision-models.pick-tool-or-skill.json, or the MCP tool get_use_case with id decision-models.pick-tool-or-skill. Other jobs in this category: /api/v1/niches?category=decision-models.
For companies
Do agents find, use and choose your tools?
An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.