Skip to content

The catalog says what this build can run

Nika v0.115.0 split the providers a build can call from those it only lists, and counted a signed-in agent app as ready only with its adapter.

The engine README today sums up Nika in three short sentences. The middle one is "Run with the model you choose." That promise is only as good as the menu. A choice is real when the tool tells you which options it can honour.

Before v0.115.0, the catalog did not. nika catalog lists the providers and models the engine knows, with their prices. The 0.115.0 notes give the numbers of that day: the catalog "ships 38 vendors and this binary carries adapters for 16". An adapter is the code inside the engine that actually talks to a provider. The other 22 vendors sat in the same list, looking the same as the 16, and no run could reach them.

The human list, the JSON output and the MCP nika_catalog tool, which coding agents read, all "rendered a reachable and an unreachable vendor identically, so the fact only arrived after the choice". You picked a provider, wrote it into a workflow, and learned at the check that this build could not call it.

v0.115.0, released on 26 August 2026, separated those two facts on every surface. Then it added a third way to reach a model.

Listing is not reaching #

The first change was a field. Every provider row now carries resolves, true when this build can call that provider. The notes stress where the value comes from: it is "derived from nika_providers::CANONICAL_IDS and never typed". In plain words, it is computed from the adapters compiled into the binary, not kept by hand where it could drift. The JSON gained the field without a format change, and the MCP tool's description now "tells an agent to pick only from rows where it is true".

The human list went through two drafts in the same release. The first marked each unreachable row with catalog only — does not resolve in this build. That was correct and hard to read: 22 of 38 rows repeated the same words, so a reader had to scan for their absence. The second draft groups instead. LOCAL and CLOUD hold only the providers this build resolves; the rest sit under one heading, CATALOG ONLY (no adapter in this build · nika check refuses these). The notes: "The group is the marker, said once."

The commands did not change; what they print did. Here are three separate commands, not one recorded session. The 0.115.0 notes name the first and the third; nika catalog --json dates from an earlier release and gained the resolves field in this one:

terminal
nika catalog
nika catalog --json
nika doctor --ping

The JSON keeps resolves on every row, because "a machine filters on resolves, it does not read headings". The notes call the third command the probe for the local engines, which the next section explains.

One word, one count #

Several smaller fixes in the release apply the same rule: a number on screen must say what it counts.

  • nika doctor, the machine report, printed « 633 models » on its pricing line, one command away from the catalog's « 69 models ». The line now counts 633 price rules and, beside them, the providers that have a price, because "several patterns can price one model, and a pattern can price models this catalog never lists".
  • A local row under a « this machine » header listed five local engines "with nothing listening on any of them". Those names were what the binary supports, not what the machine runs. The row now says how many keyless engines exist and names the probe, nika doctor --ping.
  • When a model could not be resolved, the refusal pointed at nika doctor to list the runnable providers. Measured on a fresh machine, plain nika doctor named five of the sixteen. Both refusals now point at nika catalog, and a test counts the rows the catalog actually prints.
  • The green ✔ MODELS line could be read as "the run will reach the model". It only judges whether this build can resolve the model, never whether a key is present, and now it says so: « key presence on this machine not judged (advisory: check --json access_plan · the run gate refuses NIKA-INFER-001) ».

The welcome screen also dropped an older menu, printed eleven lines below its one suggested next step, which gave a newcomer two different places to start. It now promises one first command, and the notes record the result: "43 lines down to 38."

These counts belong to v0.115.0. The catalog has changed since; read today's numbers from nika catalog itself.

A third way in: the app you already use #

The same release made agent applications a working way to reach a model in the binary people download. The engine's guide today defines the term: "an agent application you are signed into, a harness seat". Claude Code and Codex are the two the 0.115.0 notes name.

The notes say the code for it "has been on main, API-frozen, since 2026-08-08, and was compiled out of every downloadable binary". From 0.115.0 the release build includes it, and agent: tasks "can sit on a detected harness via --access <seat>". A second change let one-shot infer: tasks use a proved Codex subscription seat. The notes describe how such a call is recorded: it "records subscription quota without inventing a token count, USD price, or responding-model claim".

The first screen got the same treatment. A signed-in Claude Code used to appear as a ready seat even when its adapter, a separate program that a session actually starts, was not installed. The notes: "ready now needs three things, not two: the app is here, it is signed in, and its ACP adapter answers." ACP, the Agent Client Protocol, is the protocol Nika uses to talk to such an application. The row stays visible and names the one command that closes the gap, because "hiding an app the person has would trade one lie for another".

Last, three surfaces that each read their own source of truth about access now read one census: provider rows and harness seats joined in a single read. The run's refusals and nika explain add « or use a signed-in seat: --access … », printed only when this machine actually has one.

Why this is about choice #

In this release the model is named in the file (model:) and an agent-application path is picked on the command line (--access <seat>). Both choices belong to the person running the workflow, and both depend on an accurate menu. A list that shows paths the build cannot take does not widen the choice. It moves the failure to after the choice, when a workflow is already written around it. For an agent reading the MCP tool, it is worse: the agent acts on what the list says.

Separating "known" from "callable", and "installed" from "ready", is what keeps the choice in the user's hands.

What this does not do #

  • resolves means this build has an adapter. It does not mean the provider is reachable from your network, or that your key is valid. The MODELS line says outright that key presence is not judged there.
  • A seat is admitted on presence: the app, its sign-in, its adapter. The 0.118.0 section of the changelog, first released publicly in v0.118.6, states that "nothing yet hashes, pins or sandboxes the seat's binary", so a program that impersonates a seat on the PATH runs with your rights.
  • A one-shot call on the Codex subscription seat records quota, not a price. The run's record holds no dollar figure for it.
  • Every count in this post (38 vendors, 16 adapters, 22 catalog-only rows, 633 price rules, 10 priced providers, 69 models, five local engines) describes v0.115.0 and moves with each catalog update.

Read the v0.115.0 release, or continue with The access layer is an operating contract.