Choose a stack for an agent
stackFor 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.
Tool call template
Replace each placeholder with a value matching the input schema. Only required parameters are included; add optional parameters as needed.
{
"id": 1,
"jsonrpc": "2.0",
"method": "tools/call",
"params": {
"arguments": {
"systems": "<systems>",
"target": "<target>"
},
"name": "stack"
}
}Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| infra | No | Infrastructure the agent runs on. Only read when `actions` or `needs` imply some. | |
| model | No | 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 | No | 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 | No | 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 | No | How many users the app has or expects, when known: `approval_needed` then compares it with each vendor's number. | |
| prefer | No | 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 | Yes | 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 | No | What the agent does; default ["read"]. untrusted_input (it reads text from outside) together with write or code_execution is warned about. | |
| clients | No | 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 | No | 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 | Yes | 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 | No | 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 | No | 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 | No | 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 | No | 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. |