docs(providers): a model server you run needs no API key

[indykish]

Companion to the agentsfleet M168 provider-curation PR, which stops requiring an api_key for a local runtime.

What changed

api-reference/error-codes.mdx — UZ-PROVIDER-003's copy said an API key is required for "a named provider". A local runtime (ollama, vllm, …) is named too, so the message sent operators hunting for a key their own server never issues. It now names both exemptions. The string is kept byte-identical to error_entries.zig in the lead repo, so the two cannot drift.

fleets/model-providers.mdx — gained the section that was missing entirely: how to point at a model server you run yourself, and the two rules that differ there (no API key needed, and the model is not checked against the catalogue because the set of models you might load is yours). Two earlier claims on the same page described behavior that no longer exists and were corrected:

• "The provider id, API key, and model are all required" — the API key is now conditional.
• "A --model the provider does not serve is refused" — true for hosted providers, not for a server you run.

Verification

make lint passes — documentation check and Mintlify broken-links both green. The first draft tripped DOC-02 (31-word sentence) and DOC-10 (unexpanded LM); both are fixed rather than worked around, and the shortened UZ-PROVIDER-003 string was mirrored back into the lead repo in the same change.

🤖 Generated with Claude Code

Greptile Summary

This PR documents API-key-free credentials for locally operated model servers and updates provider error guidance. It adds the local-provider setup workflow, explains its endpoint and catalogue-validation differences, and qualifies existing credential requirements.

Confidence Score: 4/5

The documentation inconsistency should be fixed before merging because users are still promised creation-time model validation that the new local-provider flow explicitly bypasses.

The local-provider section correctly describes deferred model validation, but unchanged blanket claims in the guide introduction and create/update command reference tell users that invalid credentials cannot be stored.

Files Needing Attention: fleets/model-providers.mdx and cli/agentsfleet.mdx

Important Files Changed

Filename	Overview
fleets/model-providers.mdx	Adds the local model-server workflow, but leaves blanket catalogue-validation guarantees in this guide and the shared CLI reference inconsistent with the new exception.
api-reference/error-codes.mdx	Updates UZ-PROVIDER-003 to explain that custom endpoints and locally operated model servers may omit an API key.

Prompt To Fix All With AI

### Issue 1
fleets/model-providers.mdx:127-129
**Catalogue validation remains contradictory**

When a user creates or updates a local model-server credential, this section permits an unchecked model name while the guide introduction and shared CLI reference still guarantee catalogue validation, causing users to store misspelled models that fail only when a fleet first runs.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

_{Reviews (1): Last reviewed commit: "docs(providers): a model server you run ..." | Re-trigger Greptile}

Greptile also left 1 inline comment on this PR.

Context used:

• Knowledge Base — Model-provider credentials
• Knowledge Base — API reference documentation

[greptile-apps]

Catalogue validation remains contradictory

When a user creates or updates a local model-server credential, this section permits an unchecked model name while the guide introduction and shared CLI reference still guarantee catalogue validation, causing users to store misspelled models that fail only when a fleet first runs.

Knowledge Base Used: Model-provider credentials

Prompt To Fix With AI

This is a comment left during a code review.
Path: fleets/model-providers.mdx
Line: 127-129

Comment:
**Catalogue validation remains contradictory**

When a user creates or updates a local model-server credential, this section permits an unchecked model name while the guide introduction and shared CLI reference still guarantee catalogue validation, causing users to store misspelled models that fail only when a fleet first runs.

**Knowledge Base Used:** [Model-provider credentials](https://app.greptile.com/agentsfleet/-/custom-context/knowledge-base/agentsfleet/docs/-/docs/model-providers.md)

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

[indykish]

Withdrawing — the local-runtime carve-out this documented has been pulled out of agentsfleet#610 and deferred to its own workstream.

Both changes here describe behaviour that is no longer shipping:

• The UZ-PROVIDER-003 correction. With the keyless exemption reverted, "an API key for a named provider" is accurate again, and the string is back to matching error_entries.zig byte for byte.
• The "point at a model server you run" section, and the two claims it corrected. Both revert with it.

Reopen this branch alongside the local-runtime workstream rather than rewriting it — the prose is still right for the behaviour it describes, it just needs that behaviour to exist first.