Skip to the content

DocsBuild an agent

Stacks

POST /v1/stack: for the systems an agent works with, each system's own server or tool, the model, warnings and a config, by fixed rules.

POST /v1/stack answers a different question from /v1/resolve: not only which model, but what an agent that works with Linear, GitHub and Postgres should connect to. It names one choice a system, the model, the owners' skills, what to read first and a config to start from.

The need

FieldWhat it says
systems1 to 12 systems, by id, name or alias. Each is matched exactly against the curated list, case and punctuation aside: postgres and PostgreSQL are one system, post is none. A name that matches nothing is not an error: it comes back in unknown_systems and the rest are answered.
targetlocal: on a person's machine, with a shell. cloud: a hosted agent, which can only reach a remote server.
harnessThe client the agent runs in, by id. It decides what is possible: whether a local server can be started, whether skills load. Left out, a client that can do all of it is assumed, and assumptions says so. Unlike a system, an unknown id is a 400 whose allowed lists the ids.
tenancysingle (default) or multi, for an agent that serves several tenants.
actionsAny of read, write, code_execution and untrusted_input; ["read"] when left out.
prefercli (default) or server, for a local agent and a system that has both.
modelproviders, and in constraints any /v1/resolve parameter, as strings, except providers and the ones that ask for a team or an escalation (shape, strategy, roles, length, synthesis, origin_mix, providers_distinct): a stack has one model. A constraint /v1/resolve would refuse is a 400 here too. The model is chosen as /v1/resolve chooses one, for the purpose agent-with-tools unless you name a task or a purpose. Left out, no model is chosen and the rest is still answered.
needsWhat the agent needs beyond its systems, any of web_search, web_reading (fetching pages as text) and browser. Each adds a part to infrastructure, as code_execution among the actions adds a sandbox.
infravendors: the infrastructure vendors you have accounts with, by name. A product is chosen only among these. usage: a month of what the agent will use (seconds, vcpus, memory_gib, busy_share, searches, pages…). With it, each product is priced as a month's bill; without it, for a stated reference job. Only read when the request implies some infrastructure.
stack.sh
#!/usr/bin/env bash
# Choose a stack with curl: for an agent that works with Linear, GitHub and
# Postgres on your machine, what to connect and which model to call. Prints the
# JSON answer: `systems[].choice` is each system's server or command-line tool,
# `warnings` what to read first, and `config.mcp_json` a `.mcp.json` to start from.
#
#   RHUMB_KEY=nnm_live_... bash stack.sh
#
# RHUMB_URL overrides the base URL (default https://rhumbdata.com).
set -euo pipefail

: "${RHUMB_KEY:?Set RHUMB_KEY to your nnm_live_ key}"
RHUMB_URL="${RHUMB_URL:-https://rhumbdata.com}"

curl --silent --show-error --fail-with-body --max-time 10 \
  -X POST "$RHUMB_URL/v1/stack" \
  -H "Authorization: Bearer $RHUMB_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "systems": ["Linear", "GitHub", "Postgres"],
    "target": "local",
    "actions": ["read", "write"],
    "model": { "providers": ["gemini", "openrouter"] }
  }'
echo

The rules of a choice

For each system, only the servers its owner publishes are considered, and of those only the ones not marked deprecated. Which server is a system's own is a list we keep by hand, by the names the MCP Registry gives them; see how the directory is made. Then, in this order, the first rule that applies is the choice, and its name is the answer's rule:

RuleThe choice
cloud_remoteA cloud agent: the first of the owner's servers with a Streamable HTTP remote a client can use as it is.
local_cliA local agent, the system has a curated command-line tool, and prefer is not server: the tool. A local agent has a shell, and a tool takes no context until it is used.
local_remoteA local agent with no such tool, or one that prefers a server: the owner's remote server, if there is one the harness can reach. It works locally too, and runs none of the publisher's code on the machine.
local_packageElse the owner's server as a package pinned to one version (npm, then PyPI, then a container image), when the harness can start one.
local_cli_fallbackA local agent that preferred a server, when neither rule above found one it can use: the command-line tool after all.

Among several of an owner's servers of one kind, the one the curated list names first is taken. Where there is a first-party option that was not chosen (the server behind a command-line tool, or the tool beside a server), it is the answer's alternative.

When no rule applies the choice is none, with why and with how many servers by other publishers the directory lists:

WhyMeaning
no_official_serverIts owner lists no server of its own in the MCP Registry.
none_activeIts owner's servers are all marked deprecated.
no_remoteA cloud agent, and its owner publishes no remote server a client can use as it is.
no_usable_serverNothing of its owner's can be written into a config as it is: no usable remote, and no package pinned to a version.
harness_cannot_runIts owner publishes a server, but not of a kind the harness named can run.

Expect none often. In the catalog of 13 November 2026, 66 of the 119 curated systems had a server of their owner's, and for only 37 was it one a config can hold and a cloud agent can reach. A local agent is answered for 65 systems, 28 of them by a command-line tool; a cloud agent for 37.

Why another publisher's server is never chosen

Most servers in the registry are published by someone other than the system's owner, and some are widely used. None is ever the choice here, however popular. A server acts with your credentials for the system, and the only thing the public record can establish about a publisher is whether it is the owner. Downloads and stars measure attention, not care, and we have reviewed none of them.

So the answer is the owner's own, or nothing, with a count. The system's page lists every server for it with what is known of each, and choosing one of those is yours to weigh.

For the same reason a choice with a known problem is still the choice, with the problem said in warnings. There is no second best to fall back to, and leaving the owner's server out would hide the fact.

Warnings

Each warning is a fixed rule with an id, what it is about, a sentence, and the facts it fired on. There is no score and no risk level.

IdFires whenWhat it does not mean
untrusted_input_with_writeThe request's actions have untrusted_input together with write or code_execution: text from outside can steer an agent that can change things.It says nothing about any server. It is about what you asked the agent to do.
no_first_partyA system has no choice: its owner publishes nothing this agent can use.It does not mean nothing exists. The directory may list servers by other publishers; none is recommended here.
known_vulnerabilityAn advisory is open against the version of the chosen server's package that the registry lists.It does not mean the remote runs that version, or that the advisory applies to how you use it. No advisory does not mean none exists.
pattern_flagThe chosen server's tool definitions matched a fixed pattern rule when last read.It is a reason to read the definitions, not a finding against the server.
not_answeringThe chosen server's remote has not answered as an MCP server since the date given.It was asked from our network, with no credential. It may answer you.
tools_not_readThe chosen server's tools were never read, so what they add to a context is not known.It is not a fault. A client with no credential is not shown them, or the server was not asked at all: only a remote that declares no key is asked, never a package.
context_shareThe chosen servers' tool definitions come to over a tenth of the chosen model's context, on a harness not known to load tools only when needed.It counts only the servers whose tools were read, so the real share can be higher.
local_codeThe choice is a package the harness downloads and runs: its publisher's code runs on the machine.It does not say the code is bad. It says where it runs, and with whose access.
shared_credentialThe request says the agent serves several tenants, and the chosen server takes a fixed key or token.A server with no fixed key is not thereby safe for several tenants: how it signs each user in is yours to check.
harness_uncheckedThe choice depends on something the named harness was never checked to do.Not checked is not no. The answer assumes it can, and says so.
infra_unconfirmedA vendor you named has a figure in the answer, and its pricing page has stopped printing one of the prices the catalog holds for it.It does not mean the price is wrong, only that the nightly check could no longer find it on the page. Read the vendor's page; after a week of this the product is set aside.
infra_not_pricedA vendor you named sells the kind of infrastructure the agent needs, and has no figure here: its prices were not read, were set aside, or it does not sell what the job uses.It is not a reason to avoid the vendor. It is why that vendor was not compared or chosen.
sandbox_missingThe agent runs code, you named infrastructure vendors, and none of them sells a sandbox listed here.It does not mean you have no sandbox: the directory lists few. It says the answer could not name one among yours.

Skills

For each system given a choice, up to three of its owner's own skills, six in all, most recently changed first. They are left out for a harness known not to load skills. Nothing about a skill's content is used to pick it, and none was reviewed: read a skill before you install it, as you would any instructions you hand an agent.

The config

config.mcp_json is an object ready to save as .mcp.json, with one entry for each chosen server under the label the choice's config_key names: the id of the system it was chosen for. A server that serves two of the systems you asked for is written once, and both choices name its entry. Its warnings are given once too, about the first of the two systems: match them to a choice by about.server. The config is built only from parts checked by shape: an https address, header and variable names of a fixed form, a package name of its registry's form, pinned to one version. Nothing a publisher wrote as prose reaches it, and a server whose record does not fit gets no entry.

  • Every value you supply is the placeholder <your value>, and config.notes names each one. No key of yours is ever asked for or sent.
  • A remote that signs you in cannot be fully set up from a file: the note says the harness will open the sign-in when it first connects.
  • config.install lists the command-line tools and skills to install by hand. Nothing is installed for you.

The model

model is the single answer /v1/resolve would give for the same providers and constraints. If nothing fits, model.none says what removed the candidates, and the rest of the stack is returned all the same: the call is a 200 and is counted. A coding harness does not imply a coding model: the purpose says what the agent does, the harness only where it runs.

What it runs on

When the request implies infrastructure, the answer has an infrastructure part for each kind: a sandbox when the agent runs code, and search, reading or a browser when needs asks for them. Each part lists the products priced for it, least first, each figure with its sum, and chooses only among the vendors you name in infra.vendors. A model's cost is never added to these. See Infrastructure and prices for what a figure is for, and why no vendor is chosen for you.

Counting and caching

One counted call on every plan, with the resolve scope. The same request against the same two catalog versions (catalog_version and agent_catalog_version) on the same day gets the same answer, so cache it for a day. If the API holds no agent catalog the call answers 503 agent_catalog_unavailable, uncounted, and the rest of the API is unaffected.

Watching a stack

An answer is true on the day it is given. To be told when it changes, save the stack: on Architect an agent, sign in and press Watch under an answer, or open the answer's watch_url. You get one email on a day something in it changed, and none when nothing did. There is no API call that sets a watch. See Watching for changes.