Rhumb is an MCP server published by com.rhumbdata, hosted at rhumbdata.com. When last checked on 2026-11-22 it was working: it answers and lists its tools. It declares 9 tools, including resolve, plan, stack, platform, list_models.
Which AI model to call on your providers, and what an agent connects for each system it uses.
199 connections through 4 shared values. Tap any node.
Inferred from the host name and headers: what a server runs on, not who runs it.
Which model to call for a need, on the providers the user has accounts with: the provider's own model id, its price, capabilities and the reason. Give `providers` and a `task` or a `purpose`; add `needs`, `budget`, `origin` and other constraints to narrow it. `shape` asks for a team or an escalation ladder instead of one model. Counts as one call. An answer of `no_candidate` names what ruled every model out and the nearest miss. The answer depends on these goals; have the ones that matter in hand from the plan, the repository or the user before calling: cost (`optimize`, `max_price`, `budget`, `calls_per_user_month`); depth (`optimize`, `quality_margin`, `min_quality`, `budget`, `shape`, `needs`); maker_origin (`origin`, `origin_strict`); data_region (`regions`); languages (`languages`); data_kind (`needs`); accounts (`providers`). Send them as query parameters.
| Parameter | Type | Description |
|---|---|---|
| task | string | A task, e.g. text.chat. Required unless purpose is given. |
| hours | string | UTC hours you make your calls in, every day, as HH-HH ranges: 09-17, or 22-06 past midnight, or 00-24. A price with cheaper hours then ranks on its average over these hours instead of over a flat week. Comma-separated. |
None yet. Changes appear after a server's second daily check.
Squares are this server's values; dots are other servers sharing them. A shared address or publisher is a lead, not proof of a common operator.
| needs | string | Required capabilities, optionally with a level: "input.pdf:native". Comma-separated. |
| pages | string | PDF pages per call. A provider that prices no page bills them in input tokens, so count them in input_tokens. |
| roles | string | Team roles: "name:count" pairs, comma-separated (lead:1,workers:4), or a JSON object of role name to { count, task, needs, min_quality, optimize, distinct_vendor_from, distinct_family, judges }. Overrides and extends the strategy's roles. |
| shape | string | single (default, or the purpose's), team (with strategy or roles) or escalation (a ladder of length steps, or one model then a team with strategy or roles). |
| budget | string | Thirds of the market's prices for the task. |
| images | string | Input images per call. |
| length | string | Escalation steps, 1–5 (default 3). |
| origin | string | Model maker countries or groups: "us", "eu", "not:cn". Comma-separated. |
| exclude | string | Model ids, vendors, or provider:model pairs to leave out. Comma-separated. |
| license | string | Accepted open-weight licenses, as the catalog names them: "apache-2.0", "gemma". Comma-separated. |
| purpose | string | A purpose preset id or alias, e.g. read-receipts. |
| regions | string | Where data may be served, as region codes or their prefixes: "eu", "us", "us-east". It filters only when at least half the offerings on your providers state a region; below that, nothing is removed and the answer warns region_not_filtered. An offering that states no region is kept, with region_unsta… |
| currency | string | Adds the typical call converted to this currency, e.g. EUR. |
| optimize | string | cost (default), quality or value. |
| strategy | string | A team strategy preset: its roles and their rules. |
| languages | string | Languages your users write in, as language tags: "en", "de", "pt-BR". With one other than English, the non-English quality signal leads for a purpose that lists it; when the answer could not be compared on it, it warns language_signal_unavailable. Comma-separated. |
| max_price | string | Hard cap on the typical call, in USD, e.g. 0.002, at the day's highest rate. |
| providers * | string | Required. Provider names you have accounts with, in order of preference for ties, as /v1/vocabulary lists them; case aside, a display name or a common alias (aws, azure) names a provider too. Never keys. Comma-separated. |
| synthesis | string | Add a synthesizer role to the team. |
| origin_mix | string | Team origin rules: "eu:>=1", "cn:<=0", "us:=2". Comma-separated. |
| resolution | string | Image resolution key, e.g. "1k". |
| allow_trial | string | Allow trial offerings, which aren't for production (default false). |
| max_disk_gb | string | Local runtimes: free disk in GB for the weights. |
| max_vram_gb | string | Local runtimes: GPU memory in GB, for a full offload. On Apple silicon send the GPU's share of unified memory (macOS allows about 2/3 by default, 3/4 above 36 GB) and the total as max_memory_gb. |
| min_context | string | Minimum context window in tokens. |
| min_quality | string | Minimum quality score, 0–1. |
| allow_closed | string | Allow models the provider has closed to new accounts. Send true only if your account used them before (default false). |
| input_tokens | string | Expected input tokens per call. Each usage parameter changes that one number of the purpose's or task's typical call; what is left out stays as the preset has it. |
| service_tier | string | one of ["standard","batch","flex","priority"] |
| allow_preview | string | Allow preview models (default false). |
| audio_seconds | string | Audio seconds per call: what the model hears, or for speech.synthesize what it speaks. Priced by the second, or through the model's stated tokens a second; a model that states neither has no price for audio and is not answered. |
| max_memory_gb | string | Local runtimes: the device's memory in GB (10^9 bytes). A local offering fits when its weights, KV cache at the request's min_context (8,192 tokens when smaller) and runtime overhead fit. Cloud offerings are not limited. |
| origin_strict | string | Also require a fine-tune's base model maker to match. |
| output_images | string | Generated images per call. |
| output_tokens | string | Expected output tokens per call. |
| prompt_tokens | string | Prompt size, for long-prompt price tiers. |
| video_seconds | string | Video seconds per call: what the model makes for video.generate, else what it watches. Priced by the second, or through the model's stated tokens a second; a model that states neither has no price for video and is not answered. |
| provider_terms | string | Your terms per provider: "openrouter:byok", "gemini:free_tier", "openai:discount:0.2". Comma-separated. |
| quality_margin | string | For optimize=cost: how far below the best score still counts (default 0.05). |
| released_after | string | Only models released on or after this day, YYYY-MM-DD. |
| providers_distinct | string | Every team member, or escalation step, on a different provider. |
| cached_input_tokens | string | Input tokens served from the provider's cache. |
| calls_per_user_month | string | How many calls one user makes in a month. With it, the chosen model's price carries per_user_month and the sum behind it; without it, nothing monthly is returned. |
{
"name": "resolve",
"title": "Resolve a model",
"description": "Which model to call for a need, on the providers the user has accounts with: the provider's own model id, its price, capabilities and the reason. Give `providers` and a `task` or a `purpose`; add `needs`, `budget`, `origin` and other constraints to narrow it. `shape` asks for a team or an escalation ladder instead of one model. Counts as one call. An answer of `no_candidate` names what ruled every model out and the nearest miss. The answer depends on these goals; have the ones that matter in hand from the plan, the repository or the user before calling: cost (`optimize`, `max_price`, `budget`, `calls_per_user_month`); depth (`optimize`, `quality_margin`, `min_quality`, `budget`, `shape`, `needs`); maker_origin (`origin`, `origin_strict`); data_region (`regions`); languages (`languages`); data_kind (`needs`); accounts (`providers`). Send them as query parameters.",
"inputSchema": {
"type": "object",
"required": [
"providers"
],
"properties": {
"task": {
"type": "string",
"description": "A task, e.g. text.chat. Required unless purpose is given."
},
"hours": {
"type": "string",
"description": "UTC hours you make your calls in, every day, as HH-HH ranges: 09-17, or 22-06 past midnight, or 00-24. A price with cheaper hours then ranks on its average over these hours instead of over a flat week. Comma-separated."
},
"needs": {
"type": "string",
"description": "Required capabilities, optionally with a level: \"input.pdf:native\". Comma-separated."
},
"pages": {
"type": "string",
"pattern": "^\\d+$",
"description": "PDF pages per call. A provider that prices no page bills them in input tokens, so count them in input_tokens."
},
"roles": {
"type": "string",
"description": "Team roles: \"name:count\" pairs, comma-separated (lead:1,workers:4), or a JSON object of role name to { count, task, needs, min_quality, optimize, distinct_vendor_from, distinct_family, judges }. Overrides and extends the strategy's roles."
},
"shape": {
"enum": [
"single",
"team",
"escalation"
],
"type": "string",
"description": "single (default, or the purpose's), team (with strategy or roles) or escalation (a ladder of length steps, or one model then a team with strategy or roles)."
},
"budget": {
"enum": [
"cheap",
"medium",
"premium"
],
"type": "string",
"description": "Thirds of the market's prices for the task."
},
"images": {
"type": "string",
"pattern": "^\\d+$",
"description": "Input images per call."
},
"length": {
"type": "string",
"pattern": "^\\d+$",
"description": "Escalation steps, 1–5 (default 3)."
},
"origin": {
"type": "string",
"description": "Model maker countries or groups: \"us\", \"eu\", \"not:cn\". Comma-separated."
},
"exclude": {
"type": "string",
"description": "Model ids, vendors, or provider:model pairs to leave out. Comma-separated."
},
"license": {
"type": "string",
"description": "Accepted open-weight licenses, as the catalog names them: \"apache-2.0\", \"gemma\". Comma-separated."
},
"purpose": {
"type": "string",
"description": "A purpose preset id or alias, e.g. read-receipts."
},
"regions": {
"type": "string",
"description": "Where data may be served, as region codes or their prefixes: \"eu\", \"us\", \"us-east\". It filters only when at least half the offerings on your providers state a region; below that, nothing is removed and the answer warns region_not_filtered. An offering that states no region is kept, with region_unstated. Comma-separated."
},
"currency": {
"type": "string",
"description": "Adds the typical call converted to this currency, e.g. EUR."
},
"optimize": {
"enum": [
"cost",
"quality",
"value"
],
"type": "string",
"description": "cost (default), quality or value."
},
"strategy": {
"enum": [
"debate",
"red-team",
"competitive",
"hierarchical",
"expert-panel",
"collaborative"
],
"type": "string",
"description": "A team strategy preset: its roles and their rules."
},
"languages": {
"type": "string",
"description": "Languages your users write in, as language tags: \"en\", \"de\", \"pt-BR\". With one other than English, the non-English quality signal leads for a purpose that lists it; when the answer could not be compared on it, it warns language_signal_unavailable. Comma-separated."
},
"max_price": {
"type": "string",
"description": "Hard cap on the typical call, in USD, e.g. 0.002, at the day's highest rate."
},
"providers": {
"type": "string",
"description": "Required. Provider names you have accounts with, in order of preference for ties, as /v1/vocabulary lists them; case aside, a display name or a common alias (aws, azure) names a provider too. Never keys. Comma-separated."
},
"synthesis": {
"enum": [
"true",
"false"
],
"type": "string",
"description": "Add a synthesizer role to the team."
},
"origin_mix": {
"type": "string",
"description": "Team origin rules: \"eu:>=1\", \"cn:<=0\", \"us:=2\". Comma-separated."
},
"resolution": {
"type": "string",
"description": "Image resolution key, e.g. \"1k\"."
},
"allow_trial": {
"enum": [
"true",
"false"
],
"type": "string",
"description": "Allow trial offerings, which aren't for production (default false)."
},
"max_disk_gb": {
"type": "string",
"pattern": "^\\d+(\\.\\d+)?$",
"description": "Local runtimes: free disk in GB for the weights."
},
"max_vram_gb": {
"type": "string",
"pattern": "^\\d+(\\.\\d+)?$",
"description": "Local runtimes: GPU memory in GB, for a full offload. On Apple silicon send the GPU's share of unified memory (macOS allows about 2/3 by default, 3/4 above 36 GB) and the total as max_memory_gb."
},
"min_context": {
"type": "string",
"pattern": "^\\d+$",
"description": "Minimum context window in tokens."
},
"min_quality": {
"type": "string",
"description": "Minimum quality score, 0–1."
},
"allow_closed": {
"enum": [
"true",
"false"
],
"type": "string",
"description": "Allow models the provider has closed to new accounts. Send true only if your account used them before (default false)."
},
"input_tokens": {
"type": "string",
"pattern": "^\\d+$",
"description": "Expected input tokens per call. Each usage parameter changes that one number of the purpose's or task's typical call; what is left out stays as the preset has it."
},
"service_tier": {
"enum": [
"standard",
"batch",
"flex",
"priority"
],
"type": "string"
},
"allow_preview": {
"enum": [
"true",
"false"
],
"type": "string",
"description": "Allow preview models (default false)."
},
"audio_seconds": {
"type": "string",
"pattern": "^\\d+$",
"description": "Audio seconds per call: what the model hears, or for speech.synthesize what it speaks. Priced by the second, or through the model's stated tokens a second; a model that states neither has no price for audio and is not answered."
},
"max_memory_gb": {
"type": "string",
"pattern": "^\\d+(\\.\\d+)?$",
"description": "Local runtimes: the device's memory in GB (10^9 bytes). A local offering fits when its weights, KV cache at the request's min_context (8,192 tokens when smaller) and runtime overhead fit. Cloud offerings are not limited."
},
"origin_strict": {
"enum": [
"true",
"false"
],
"type": "string",
"description": "Also require a fine-tune's base model maker to match."
},
"output_images": {
"type": "string",
"pattern": "^\\d+$",
"description": "Generated images per call."
},
"output_tokens": {
"type": "string",
"pattern": "^\\d+$",
"description": "Expected output tokens per call."
},
"prompt_tokens": {
"type": "string",
"pattern": "^\\d+$",
"description": "Prompt size, for long-prompt price tiers."
},
"video_seconds": {
"type": "string",
"pattern": "^\\d+$",
"description": "Video seconds per call: what the model makes for video.generate, else what it watches. Priced by the second, or through the model's stated tokens a second; a model that states neither has no price for video and is not answered."
},
"provider_terms": {
"type": "string",
"description": "Your terms per provider: \"openrouter:byok\", \"gemini:free_tier\", \"openai:discount:0.2\". Comma-separated."
},
"quality_margin": {
"type": "string",
"description": "For optimize=cost: how far below the best score still counts (default 0.05)."
},
"released_after": {
"type": "string",
"description": "Only models released on or after this day, YYYY-MM-DD."
},
"providers_distinct": {
"enum": [
"true",
"false"
],
"type": "string",
"description": "Every team member, or escalation step, on a different provider."
},
"cached_input_tokens": {
"type": "string",
"pattern": "^\\d+$",
"description": "Input tokens served from the provider's cache."
},
"calls_per_user_month": {
"type": "string",
"pattern": "^\\d+$",
"description": "How many calls one user makes in a month. With it, the chosen model's price carries per_user_month and the sum behind it; without it, nothing monthly is returned."
}
},
"additionalProperties": false
},
"annotations": {
"readOnlyHint": true,
"openWorldHint": false
}
}Reads a description of the work (a prompt, tags, attachment types and sizes) into a need, then resolves it as `resolve` does. Use `resolve` when you can state the task or purpose yourself; use this when you only have the user's words. Counts as more than one call. The answer depends on these goals; have the ones that matter in hand from the plan, the repository or the user before calling: cost (`optimize`, `max_price`, `budget`, `calls_per_user_month`); depth (`optimize`, `quality_margin`, `min_quality`, `budget`, `shape`, `needs`); maker_origin (`origin`, `origin_strict`); data_region (`regions`); languages (`languages`); data_kind (`needs`); accounts (`providers`). Send them as strings in `constraints`; only `providers` and `shape` go at the top level.
| Parameter | Type | Description |
|---|---|---|
| tags | array | Free tags, e.g. ["vision", "read receipts", "usa", "medium"]. |
| shape | string | As in /v1/resolve, or auto: a team purpose's team and an escalation purpose's ladder; else a team when the reading says several models would help (ensemble_helps), an escalation when the request gives length, else single. |
| device | object | The device's memory, GPU memory and free disk in GB, for local runtimes: facts that become max_memory_gb, max_vram_gb and max_disk_gb unless constraints set them. |
| prompt | string | The request to read. Sent only to the decision model you choose; never stored or logged. |
| decision | object | |
| providers * | array | Required. Provider names, in order of preference for ties. |
| attachments | array | |
| constraints | object | Any /v1/resolve query parameter except providers, as strings; these win over the reading. When these and the recognised tags settle the task or purpose and the budget, give needs (even empty), no tag is unrecognised, and no shape is left to the reading, the prompt is not read: the reading is empty a… |
{
"name": "plan",
"title": "Plan from a description",
"description": "Reads a description of the work (a prompt, tags, attachment types and sizes) into a need, then resolves it as `resolve` does. Use `resolve` when you can state the task or purpose yourself; use this when you only have the user's words. Counts as more than one call. The answer depends on these goals; have the ones that matter in hand from the plan, the repository or the user before calling: cost (`optimize`, `max_price`, `budget`, `calls_per_user_month`); depth (`optimize`, `quality_margin`, `min_quality`, `budget`, `shape`, `needs`); maker_origin (`origin`, `origin_strict`); data_region (`regions`); languages (`languages`); data_kind (`needs`); accounts (`providers`). Send them as strings in `constraints`; only `providers` and `shape` go at the top level.",
"inputSchema": {
"type": "object",
"required": [
"providers"
],
"properties": {
"tags": {
"type": "array",
"items": {
"type": "string",
"maxLength": 100
},
"maxItems": 50,
"description": "Free tags, e.g. [\"vision\", \"read receipts\", \"usa\", \"medium\"]."
},
"shape": {
"enum": [
"single",
"team",
"escalation",
"auto"
],
"type": "string",
"description": "As in /v1/resolve, or auto: a team purpose's team and an escalation purpose's ladder; else a team when the reading says several models would help (ensemble_helps), an escalation when the request gives length, else single."
},
"device": {
"type": "object",
"properties": {
"disk_gb": {
"type": "number",
"exclusiveMinimum": 0
},
"vram_gb": {
"type": "number",
"exclusiveMinimum": 0
},
"memory_gb": {
"type": "number",
"exclusiveMinimum": 0
}
},
"description": "The device's memory, GPU memory and free disk in GB, for local runtimes: facts that become max_memory_gb, max_vram_gb and max_disk_gb unless constraints set them.",
"additionalProperties": false
},
"prompt": {
"type": "string",
"description": "The request to read. Sent only to the decision model you choose; never stored or logged."
},
"decision": {
"type": "object",
"properties": {
"model": {
"enum": [
"openai",
"local",
"jev"
],
"type": "string",
"description": "openai (default) or local; jev is for internal accounts only."
},
"allow_remote": {
"type": "boolean",
"description": "false keeps the prompt on the server: the local classifier only. When the account turns on \"Keep prompts on Rhumb's server\" in Settings, every request reads this way."
}
},
"additionalProperties": false
},
"providers": {
"type": "array",
"items": {
"type": "string"
},
"minItems": 1,
"description": "Required. Provider names, in order of preference for ties."
},
"attachments": {
"type": "array",
"items": {
"type": "object",
"required": [
"type",
"bytes"
],
"properties": {
"type": {
"type": "string",
"description": "MIME type, e.g. image/jpeg. Only type and size are sent, never the file."
},
"bytes": {
"type": "integer",
"maximum": 9007199254740991,
"minimum": 0
},
"pages": {
"type": "integer",
"maximum": 9007199254740991,
"exclusiveMinimum": 0
},
"seconds": {
"type": "number",
"minimum": 0
}
},
"additionalProperties": false
},
"maxItems": 20
},
"constraints": {
"type": "object",
"description": "Any /v1/resolve query parameter except providers, as strings; these win over the reading. When these and the recognised tags settle the task or purpose and the budget, give needs (even empty), no tag is unrecognised, and no shape is left to the reading, the prompt is not read: the reading is empty and decision.model is local.",
"propertyNames": {
"type": "string"
},
"additionalProperties": {
"type": "string"
}
}
},
"additionalProperties": false
},
"annotations": {
"readOnlyHint": true,
"openWorldHint": false
}
}For an agent that has to work with named systems (Linear, GitHub, Postgres…): what to connect for each, and which model to call. Give `systems`, `target` (local or cloud) and the `harness` the agent runs in; add `model.providers` to get a model too. Each system's choice is its owner's own MCP server or command-line tool, or none, picked by fixed rules from public records and not reviewed; `warnings` say what to read first, and `config.mcp_json` is a config to start from, with placeholders for every value the user supplies. A name that matches no system comes back in `unknown_systems`. Counts as one call. The answer depends on these goals; have the ones that matter in hand from the plan, the repository or the user before calling: cost (`optimize`, `max_price`, `budget`, `calls_per_user_month`); maker_origin (`origin`, `origin_strict`); users_in (`users_in`); data_kind (`needs`); scale (`scale`); accounts (`providers`, `infra.vendors`); clients (`clients`). On `stack`, the app's goals go by their own names at the top (`data_kind`, `users_in`, `scale`, `clients`), a model's parameters in `model.constraints` and providers in `model.providers`; infrastructure vendors are `infra.vendors`. The stack's own top-level `needs` are what the agent does (`web_search`, `web_reading`, `browser`), never a model's.
| Parameter | Type | Description |
|---|---|---|
| infra | object | Infrastructure the agent runs on. Only read when `actions` or `needs` imply some. |
| model | object | The providers to choose the agent's model on, as /v1/resolve would. Left out, no model is chosen and the rest is still answered. |
| needs | array | What the agent needs beyond its systems: web_search, web_reading (fetching pages as text), browser. Each adds a part to `infrastructure`, as `code_execution` in `actions` adds a sandbox. |
| scale | string | What stage the app is at: prototype (for yourself), launch (to the public) or live. With launch or live, a system that admits only a few users before its vendor reviews the app is warned about (`approval_needed`). |
| users | integer | How many users the app has or expects, when known: `approval_needed` then compares it with each vendor's number. |
| prefer | string | For a local agent and a system with a curated command-line tool: cli (default) chooses the tool, server its owner's server when there is one. |
| target * | string | Required. Where the agent runs. local: on a person's machine, with a shell, so a system's command-line tool can be the choice. cloud: only a remote server qualifies. |
| actions | array | What the agent does; default ["read"]. untrusted_input (it reads text from outside) together with write or code_execution is warned about. |
| clients | array | Where the app runs for its users. A system read only on a device (Apple Health on iOS, Health Connect on Android) with none of its devices here is warned about (`on_device_only`). |
| harness | string | The client the agent runs in, by id (the 400 for an unknown one lists them). It decides what is possible: a local server needs one that can start it, skills one that loads them. Left out, a client that can do all of it is assumed, and `assumptions` says so. |
| systems * | array | Required. The systems the agent works with, by id, name or alias, e.g. ["Linear", "github", "postgres"]. Each is matched exactly (case and punctuation aside) against the curated list; one that matches nothing comes back in `unknown_systems` and the rest are answered. |
| tenancy | string | single (default): the agent acts for one user or team. multi: it serves several tenants, so a server that takes one fixed key is warned about. |
| building | array | What is being built, as skill tags. An owner's skills whose read carries more of them come first; a platform's owner's general skills are offered only when their read carries one. A skill not yet read keeps its place by date. |
| users_in | string | The goal users_in: where the app's users are (worldwide, eu: the EU among them, us: the US only). Recorded and listed in `assumed` when left out; no rule reads it yet, so it never changes the answer and is never asked. |
| data_kind | string | The goal data_kind. health, financial or hipaa adds zdr and no_training to the model's needs (when `model` is given), as /v1/resolve takes them; none adds nothing. |
{
"name": "stack",
"title": "Choose a stack for an agent",
"description": "For an agent that has to work with named systems (Linear, GitHub, Postgres…): what to connect for each, and which model to call. Give `systems`, `target` (local or cloud) and the `harness` the agent runs in; add `model.providers` to get a model too. Each system's choice is its owner's own MCP server or command-line tool, or none, picked by fixed rules from public records and not reviewed; `warnings` say what to read first, and `config.mcp_json` is a config to start from, with placeholders for every value the user supplies. A name that matches no system comes back in `unknown_systems`. Counts as one call. The answer depends on these goals; have the ones that matter in hand from the plan, the repository or the user before calling: cost (`optimize`, `max_price`, `budget`, `calls_per_user_month`); maker_origin (`origin`, `origin_strict`); users_in (`users_in`); data_kind (`needs`); scale (`scale`); accounts (`providers`, `infra.vendors`); clients (`clients`). On `stack`, the app's goals go by their own names at the top (`data_kind`, `users_in`, `scale`, `clients`), a model's parameters in `model.constraints` and providers in `model.providers`; infrastructure vendors are `infra.vendors`. The stack's own top-level `needs` are what the agent does (`web_search`, `web_reading`, `browser`), never a model's.",
"inputSchema": {
"type": "object",
"required": [
"systems",
"target"
],
"properties": {
"infra": {
"type": "object",
"properties": {
"usage": {
"type": "object",
"properties": {
"runs": {
"type": "number",
"maximum": 1000000000000,
"minimum": 0
},
"pages": {
"type": "number",
"maximum": 1000000000000,
"minimum": 0
},
"vcpus": {
"type": "number",
"maximum": 1000000000000,
"minimum": 0
},
"seconds": {
"type": "number",
"maximum": 1000000000000,
"minimum": 0,
"description": "How long a sandbox or a browser exists in a month, in all: seconds on the wall clock. Given with `vcpus` and `memory_gib` for an agent that runs code, it is the sandbox's: a browser's hours are then not stated, and it is priced for its reference job."
},
"searches": {
"type": "number",
"maximum": 1000000000000,
"minimum": 0
},
"busy_share": {
"type": "number",
"maximum": 1,
"minimum": 0,
"description": "The share of that time its processors compute, 0 to 1; 1 when left out."
},
"memory_gib": {
"type": "number",
"maximum": 1000000000000,
"minimum": 0
},
"transfer_gib": {
"type": "number",
"maximum": 1000000000000,
"minimum": 0
},
"results_per_search": {
"type": "number",
"maximum": 1000000000000,
"minimum": 0
}
},
"description": "A month of what the agent will use. With it, each product is priced as a month's bill on its cheapest plan; without it, for a stated reference job at listed rates, which the answer names. Nothing is assumed about your usage.",
"additionalProperties": false
},
"vendors": {
"type": "array",
"items": {
"type": "string",
"pattern": "\\S",
"maxLength": 100
},
"maxItems": 12,
"description": "The infrastructure vendors you have accounts with, by name, e.g. [\"Vercel\", \"e2b\"]. A product is chosen only among these, by price alone; with none named, products are listed and none is chosen. A vendor's name is every product of that vendor's; a product's own id, written exactly, is that product alone. A name that matches nothing comes back in `unknown_infra_vendors`."
}
},
"description": "Infrastructure the agent runs on. Only read when `actions` or `needs` imply some.",
"additionalProperties": false
},
"model": {
"type": "object",
"required": [
"providers"
],
"properties": {
"providers": {
"type": "array",
"items": {
"type": "string"
},
"minItems": 1,
"description": "Provider names, in order of preference for ties."
},
"constraints": {
"type": "object",
"description": "Any /v1/resolve query parameter except providers, as strings, and except shape, strategy, roles, length, synthesis, origin_mix, providers_distinct: a stack has one model. With neither task nor purpose, the purpose is agent-with-tools.",
"propertyNames": {
"type": "string"
},
"additionalProperties": {
"type": "string"
}
}
},
"description": "The providers to choose the agent's model on, as /v1/resolve would. Left out, no model is chosen and the rest is still answered.",
"additionalProperties": false
},
"needs": {
"type": "array",
"items": {
"enum": [
"web_search",
"web_reading",
"browser"
],
"type": "string"
},
"description": "What the agent needs beyond its systems: web_search, web_reading (fetching pages as text), browser. Each adds a part to `infrastructure`, as `code_execution` in `actions` adds a sandbox."
},
"scale": {
"enum": [
"prototype",
"launch",
"live"
],
"type": "string",
"description": "What stage the app is at: prototype (for yourself), launch (to the public) or live. With launch or live, a system that admits only a few users before its vendor reviews the app is warned about (`approval_needed`)."
},
"users": {
"type": "integer",
"maximum": 1000000000,
"description": "How many users the app has or expects, when known: `approval_needed` then compares it with each vendor's number.",
"exclusiveMinimum": 0
},
"prefer": {
"enum": [
"cli",
"server"
],
"type": "string",
"description": "For a local agent and a system with a curated command-line tool: cli (default) chooses the tool, server its owner's server when there is one."
},
"target": {
"enum": [
"local",
"cloud"
],
"type": "string",
"description": "Required. Where the agent runs. local: on a person's machine, with a shell, so a system's command-line tool can be the choice. cloud: only a remote server qualifies."
},
"actions": {
"type": "array",
"items": {
"enum": [
"read",
"write",
"code_execution",
"untrusted_input"
],
"type": "string"
},
"minItems": 1,
"description": "What the agent does; default [\"read\"]. untrusted_input (it reads text from outside) together with write or code_execution is warned about."
},
"clients": {
"type": "array",
"items": {
"enum": [
"web",
"ios",
"android",
"desktop",
"server"
],
"type": "string"
},
"minItems": 1,
"description": "Where the app runs for its users. A system read only on a device (Apple Health on iOS, Health Connect on Android) with none of its devices here is warned about (`on_device_only`)."
},
"harness": {
"type": "string",
"pattern": "\\S",
"maxLength": 100,
"description": "The client the agent runs in, by id (the 400 for an unknown one lists them). It decides what is possible: a local server needs one that can start it, skills one that loads them. Left out, a client that can do all of it is assumed, and `assumptions` says so."
},
"systems": {
"type": "array",
"items": {
"type": "string",
"pattern": "\\S",
"maxLength": 100
},
"maxItems": 24,
"minItems": 1,
"description": "Required. The systems the agent works with, by id, name or alias, e.g. [\"Linear\", \"github\", \"postgres\"]. Each is matched exactly (case and punctuation aside) against the curated list; one that matches nothing comes back in `unknown_systems` and the rest are answered."
},
"tenancy": {
"enum": [
"single",
"multi"
],
"type": "string",
"description": "single (default): the agent acts for one user or team. multi: it serves several tenants, so a server that takes one fixed key is warned about."
},
"building": {
"type": "array",
"items": {
"enum": [
"api",
"auth",
"automation",
"billing",
"browser",
"charts",
"ci",
"cli",
"cloud",
"code-review",
"containers",
"databases",
"debugging",
"deployment",
"docs",
"email",
"evals",
"files",
"frontend",
"git",
"images",
"infrastructure",
"kubernetes",
"logging",
"mcp",
"migration",
"ml",
"mobile",
"monitoring",
"networking",
"notebooks",
"payments",
"pdf",
"performance",
"presentations",
"prompts",
"refactoring",
"scraping",
"search",
"serverless",
"spreadsheets",
"sql",
"storage",
"troubleshooting",
"video",
"web"
],
"type": "string"
},
"maxItems": 8,
"minItems": 1,
"description": "What is being built, as skill tags. An owner's skills whose read carries more of them come first; a platform's owner's general skills are offered only when their read carries one. A skill not yet read keeps its place by date."
},
"users_in": {
"enum": [
"worldwide",
"eu",
"us"
],
"type": "string",
"description": "The goal users_in: where the app's users are (worldwide, eu: the EU among them, us: the US only). Recorded and listed in `assumed` when left out; no rule reads it yet, so it never changes the answer and is never asked."
},
"data_kind": {
"enum": [
"none",
"health",
"financial",
"hipaa"
],
"type": "string",
"description": "The goal data_kind. health, financial or hipaa adds zdr and no_training to the model's needs (when `model` is given), as /v1/resolve takes them; none adds nothing."
}
},
"additionalProperties": false
},
"annotations": {
"readOnlyHint": true,
"openWorldHint": false
}
}Where an app's agent runs: give `requires` (baa, gpu, private_network, edge, long_running, websockets, scheduled_jobs, large_uploads) with `longest_turn_seconds` and `largest_upload_mb` where they apply, and the goals you have. The answer is `major` (a platform is warranted: a requirement no curated runtime meets, a model provider it hosts, or an account held; name its `system` in a `stack` call next) or `generic` (any host meeting `host_must` will do, with the runtimes that meet it priced for a stated job). Every host ruled out comes with its own page's words; a host whose pages do not state a fact is neither ruled in nor out. A choice is made only when the rules settle it; otherwise it is the user's. Counts as one call. The answer depends on these goals; have the ones that matter in hand from the plan, the repository or the user before calling: users_in (`users_in`); data_region (`data_region`); data_kind (`data_kind`); scale (`scale`); accounts (`accounts`, `providers`); clients (`clients`). On `platform`, a goal is sent by its own name (`data_kind`, `data_region`, `users_in`, `scale`, `clients`, `accounts`) with the values that call lists, and the model providers as `providers`; `needs`, `regions` and `infra.vendors` are not its parameters.
| Parameter | Type | Description |
|---|---|---|
| scale | string | The goal scale. With launch or live, a free tier that stops serving when idle is warned of. |
| clients | array | The goal clients: where the app runs for its users. |
| accounts | array | The goal accounts: platforms (gcp, aws, azure, cloudflare), infrastructure products or vendors the user holds an account with. A platform held is warranted; a choice among several is made only among those held. |
| requires | array | What the app needs of a host. long_running needs longest_turn_seconds and large_uploads needs largest_upload_mb; either number given alone adds its requirement. |
| users_in | string | The goal users_in: where the app's users are (worldwide, eu: the EU among them, us: the US only). Recorded and listed in `assumed` when left out; no rule reads it yet, so it never changes the answer and is never asked. |
| data_kind | string | The goal data_kind. hipaa (for a covered entity) requires a business associate agreement; any but none, with ios among clients, adds Apple's disclosure rule. |
| providers | array | The model providers the app calls, as sent to /v1/resolve. A platform that hosts one is warranted. |
| data_region | string | The goal data_region. No host's region is curated yet, so it is warned of, never filtered on. |
| largest_upload_mb | number | The largest file one request carries, in MB: a host whose request body is smaller is ruled out, and the answer warns to upload by a signed address. |
| longest_turn_seconds | integer | The longest one request or run may take, in seconds: a host that stops one sooner is ruled out. |
{
"name": "platform",
"title": "Choose where an app's agent runs",
"description": "Where an app's agent runs: give `requires` (baa, gpu, private_network, edge, long_running, websockets, scheduled_jobs, large_uploads) with `longest_turn_seconds` and `largest_upload_mb` where they apply, and the goals you have. The answer is `major` (a platform is warranted: a requirement no curated runtime meets, a model provider it hosts, or an account held; name its `system` in a `stack` call next) or `generic` (any host meeting `host_must` will do, with the runtimes that meet it priced for a stated job). Every host ruled out comes with its own page's words; a host whose pages do not state a fact is neither ruled in nor out. A choice is made only when the rules settle it; otherwise it is the user's. Counts as one call. The answer depends on these goals; have the ones that matter in hand from the plan, the repository or the user before calling: users_in (`users_in`); data_region (`data_region`); data_kind (`data_kind`); scale (`scale`); accounts (`accounts`, `providers`); clients (`clients`). On `platform`, a goal is sent by its own name (`data_kind`, `data_region`, `users_in`, `scale`, `clients`, `accounts`) with the values that call lists, and the model providers as `providers`; `needs`, `regions` and `infra.vendors` are not its parameters.",
"inputSchema": {
"type": "object",
"properties": {
"scale": {
"enum": [
"prototype",
"launch",
"live"
],
"type": "string",
"description": "The goal scale. With launch or live, a free tier that stops serving when idle is warned of."
},
"clients": {
"type": "array",
"items": {
"enum": [
"web",
"ios",
"android",
"desktop",
"server"
],
"type": "string"
},
"minItems": 1,
"description": "The goal clients: where the app runs for its users."
},
"accounts": {
"type": "array",
"items": {
"type": "string",
"maxLength": 80,
"minLength": 1
},
"maxItems": 12,
"description": "The goal accounts: platforms (gcp, aws, azure, cloudflare), infrastructure products or vendors the user holds an account with. A platform held is warranted; a choice among several is made only among those held."
},
"requires": {
"type": "array",
"items": {
"enum": [
"baa",
"gpu",
"private_network",
"edge",
"long_running",
"websockets",
"scheduled_jobs",
"large_uploads"
],
"type": "string"
},
"description": "What the app needs of a host. long_running needs longest_turn_seconds and large_uploads needs largest_upload_mb; either number given alone adds its requirement."
},
"users_in": {
"enum": [
"worldwide",
"eu",
"us"
],
"type": "string",
"description": "The goal users_in: where the app's users are (worldwide, eu: the EU among them, us: the US only). Recorded and listed in `assumed` when left out; no rule reads it yet, so it never changes the answer and is never asked."
},
"data_kind": {
"enum": [
"none",
"health",
"financial",
"hipaa"
],
"type": "string",
"description": "The goal data_kind. hipaa (for a covered entity) requires a business associate agreement; any but none, with ios among clients, adds Apple's disclosure rule."
},
"providers": {
"type": "array",
"items": {
"type": "string"
},
"description": "The model providers the app calls, as sent to /v1/resolve. A platform that hosts one is warranted."
},
"data_region": {
"enum": [
"eu",
"us"
],
"type": "string",
"description": "The goal data_region. No host's region is curated yet, so it is warned of, never filtered on."
},
"largest_upload_mb": {
"type": "number",
"maximum": 100000,
"description": "The largest file one request carries, in MB: a host whose request body is smaller is ruled out, and the answer warns to upload by a signed address.",
"exclusiveMinimum": 0
},
"longest_turn_seconds": {
"type": "integer",
"maximum": 86400,
"description": "The longest one request or run may take, in seconds: a host that stops one sooner is ruled out.",
"exclusiveMinimum": 0
}
},
"additionalProperties": false
},
"annotations": {
"readOnlyHint": true,
"openWorldHint": false
}
}The catalog's models with each provider's offering, price, capabilities and quality scores, filtered by task, providers and needs. Returns 20 models unless `limit` says otherwise; page with `offset`. A full record is large: to see which ids a provider lists, which are deprecated or ending and what they cost, send `view: "compact"` (each model's id, release date and one quality score, each offering's id, status, end dates, per_1m rates and capability names), with a `limit` up to 500 to take a provider's whole list in one page; then `get_model` or the full view for the models you need. Counts as one call.
| Parameter | Type | Description |
|---|---|---|
| task | string | Only models for this task. |
| view | string | full (default): every field of every model. compact: each model's id, release date and one quality score, and each offering's id, status, end dates, current per_1m rates and capability names, about a fifth of the size; to scan which ids a provider lists, which are deprecated and what they cost befor… |
| limit | string | Page size, 1–500 (default 20). |
| needs | string | Only offerings with these capabilities. Comma-separated. |
| offset | string | Models to skip (default 0). |
| providers | string | Only offerings on these providers. Comma-separated. |
| include_hidden | string | Include offerings that are never recommended (default false). |
| include_retired | string | Include retired offerings (default false). |
{
"name": "list_models",
"title": "List models",
"description": "The catalog's models with each provider's offering, price, capabilities and quality scores, filtered by task, providers and needs. Returns 20 models unless `limit` says otherwise; page with `offset`. A full record is large: to see which ids a provider lists, which are deprecated or ending and what they cost, send `view: \"compact\"` (each model's id, release date and one quality score, each offering's id, status, end dates, per_1m rates and capability names), with a `limit` up to 500 to take a provider's whole list in one page; then `get_model` or the full view for the models you need. Counts as one call.",
"inputSchema": {
"type": "object",
"properties": {
"task": {
"type": "string",
"description": "Only models for this task."
},
"view": {
"enum": [
"full",
"compact"
],
"type": "string",
"description": "full (default): every field of every model. compact: each model's id, release date and one quality score, and each offering's id, status, end dates, current per_1m rates and capability names, about a fifth of the size; to scan which ids a provider lists, which are deprecated and what they cost before asking for full records."
},
"limit": {
"type": "string",
"pattern": "^\\d+$",
"description": "Page size, 1–500 (default 20)."
},
"needs": {
"type": "string",
"description": "Only offerings with these capabilities. Comma-separated."
},
"offset": {
"type": "string",
"pattern": "^\\d+$",
"description": "Models to skip (default 0)."
},
"providers": {
"type": "string",
"description": "Only offerings on these providers. Comma-separated."
},
"include_hidden": {
"enum": [
"true",
"false"
],
"type": "string",
"description": "Include offerings that are never recommended (default false)."
},
"include_retired": {
"enum": [
"true",
"false"
],
"type": "string",
"description": "Include retired offerings (default false)."
}
},
"additionalProperties": false
},
"annotations": {
"readOnlyHint": true,
"openWorldHint": false
}
}One model by id (`vendor/model`, e.g. `anthropic/claude-sonnet-4.5`): every provider's offering with its price history, capabilities, limits and dates, and the model's quality scores. Counts as one call.
| Parameter | Type | Description |
|---|---|---|
| at | string | ISO date or time: return the price revision in force then. |
| id * | string | The model's id, `vendor/model`, or an alias of it. |
| provider | string | Only this provider's offering. |
{
"name": "get_model",
"title": "Get one model",
"description": "One model by id (`vendor/model`, e.g. `anthropic/claude-sonnet-4.5`): every provider's offering with its price history, capabilities, limits and dates, and the model's quality scores. Counts as one call.",
"inputSchema": {
"type": "object",
"required": [
"id"
],
"properties": {
"at": {
"type": "string",
"description": "ISO date or time: return the price revision in force then."
},
"id": {
"type": "string",
"description": "The model's id, `vendor/model`, or an alias of it."
},
"provider": {
"type": "string",
"description": "Only this provider's offering."
}
},
"additionalProperties": false
},
"annotations": {
"readOnlyHint": true,
"openWorldHint": false
}
}The purpose presets `resolve` accepts: named jobs such as reading receipts or a coding agent, each with the task, needs and budget it stands for. Free.
No parameters.
{
"name": "list_purposes",
"title": "List purposes",
"description": "The purpose presets `resolve` accepts: named jobs such as reading receipts or a coding agent, each with the task, needs and budget it stands for. Free.",
"inputSchema": {
"type": "object",
"properties": {},
"additionalProperties": false
},
"annotations": {
"readOnlyHint": true,
"openWorldHint": false
}
}The goals an answer depends on (cost, depth, where a model's maker is based, where data may stay, languages, the kind of data, the accounts already held), each with one question to ask the user, the options to offer and the parameters each answer is sent as, and the rules for asking. Call it once when planning an app. Free.
No parameters.
{
"name": "list_goals",
"title": "List the goals",
"description": "The goals an answer depends on (cost, depth, where a model's maker is based, where data may stay, languages, the kind of data, the accounts already held), each with one question to ask the user, the options to offer and the parameters each answer is sent as, and the rules for asking. Call it once when planning an app. Free.",
"inputSchema": {
"type": "object",
"properties": {},
"additionalProperties": false
},
"annotations": {
"readOnlyHint": true,
"openWorldHint": false
}
}Every name the other tools accept: tasks, capabilities and their levels, providers, origin groups, budgets and team strategies. Free.
No parameters.
{
"name": "get_vocabulary",
"title": "Get the vocabulary",
"description": "Every name the other tools accept: tasks, capabilities and their levels, providers, origin groups, budgets and team strategies. Free.",
"inputSchema": {
"type": "object",
"properties": {},
"additionalProperties": false
},
"annotations": {
"readOnlyHint": true,
"openWorldHint": false
}
}Rhumb answers which model to use on the providers the user has accounts with. Call `resolve` with those providers and a task or purpose before choosing a model for the user's code; never guess a provider's model id or price. For an agent that has to work with named systems, call `stack` with the systems, where the agent runs and its harness: it answers each system's own MCP server or command-line tool by fixed rules over public records, unreviewed, with warnings to show the user. When its answer has `watch_url`, offer that address to the user as a link (they can be emailed when the stack changes); do not open it, and never say a watch was set. Send provider and system names only, never the user's provider or system keys, prompts' contents or files. While the user's agent plans an app, Rhumb is a helper at its decisions, not the planner: every answer depends on goals (cost, depth, maker_origin, users_in, data_region, languages, data_kind, scale, accounts, clients), which each tool's description names with the parameters they are sent as. Ask only the goals the next call depends on, one at a time, and at most five in a planning session; state the rest as assumptions when showing the answer. Look first: the plan or spec already written, then the repository (dependency, environment variable and config file names, never values), then the conversation. Ask the user only what none of these answers. Rhumb keeps no latency data, so speed is never asked.
Names, descriptions and schemas were written by the server and are shown as data.
Shared addresses on big platforms (Cloudflare, Vercel and the like) put unrelated servers together; those are marked shared_infrastructure. A match is a pivot, not a link between operators.