OmniRoute is having its breakout month. OmniRoute is a free, MIT-licensed AI gateway that puts one OpenAI-compatible endpoint in front of many model providers (352 of them, by its own README). It sits at 60,112 GitHub stars, and "omniroute" is a Breakout riser in Google Trends as of September 2, 2026. So we did what we do for skills with momentum: we read the repo's designated start-here skill for REST access, omni-auth, byte by byte at commit 93265ee, and published the full walkthrough as a line-by-line deep dive on the omni-auth listing.
The file is 7,168 bytes over 150 lines, and the most interesting thing in it is a table that has quietly died.
The file tells you it was not written by a person
The first line after the frontmatter:
<!-- generated by src/lib/agentSkills/generator.ts; manual edits will be overwritten -->
OmniRoute generates its skill files from the application's API surface, the way OpenAPI docs come out of route definitions. The endpoints in the file cannot drift from the running server. The parts the generator does not touch, we found out, can.
What the skill actually teaches an agent
Most agent skills are prose instructions; this one is closer to a compressed API reference. It teaches three things, and they are the right three:
-
Two credential types, and the boundary between them. API calls take a Bearer key. The dashboard takes a password and returns a session cookie. The file says this in two sentences, and an agent that mixes them up gets a clean
401instead of a mystery. - A two-variable setup. There is no Claude Code section or Cursor section because none is needed: any agent that speaks OpenAI-style REST points at the same base URL. Running the server itself belongs to the sibling cli-serve skill.
export OMNIROUTE_URL=http://localhost:20128
export OMNIROUTE_KEY=... # minted in the dashboard
curl $OMNIROUTE_URL/api/health # {"ok":true}
-
A status-code-to-recovery table.
401means refresh your key,503 Provider circuit openmeans waitRetry-Afterand let auto-fallback work. Four bullets that stop an agent from retry-hammering a provider whose circuit breaker is open.
For a generated file, the OIDC section is unusually explicit: state-cookie CSRF protection, ID-token verification against the issuer's JWKS, an optional subject allowlist, and a 30-day session JWT. That paragraph is where the dashboard's security posture actually lives, and you can read it in the skill file instead of the source.
The dead tables
The back half of the file is older, and it says so: a comment marks it as content migrated from an earlier skill and preserved. It ends with the skill's most useful-looking artifact: two tables linking 17 sibling skills by raw URL. Chat, images, TTS, embeddings, routing, compression, the MCP server, the CLI family.
We fetched all 17 URLs on September 2, 2026. Sixteen return 404. The only one that resolves is omni-mcp, the only entry whose name survived the catalog's rename to the current omni-* and cli-* families. The repo holds 46 SKILL.md files on disk; its own live skills index lists 45 of them correctly, one directory up from the dead tables.
The defense is printed in the file itself: the block is labeled as preserved legacy content, and the generator's notice says manual edits get overwritten, so a hand fix would not stick. The consequence still lands on the agent, though. An assistant that follows the start-here skill's own index to load a capability fetches sixteen 404 pages while the working index sits beside it.
The preserved block has a second tell: it says 327 providers while the README at the same commit says "352 providers" four times. Content outside the generator's reach trails the product. Anything below that migration comment deserves a date check.
Is it safe to route your prompts through it?
Yes, if you keep it on your own machine. A gateway sees every prompt that passes through it, so the trust question is about where it runs. OmniRoute is self-hosted by default, MIT-licensed end to end, with password plus optional OIDC-allowlist login and CSRF-guarded flows. The 30-day session token is long-lived, and the allowlist only protects you if you configure it. For calibration on how rare any scrutiny is in this ecosystem: our agent skill security census, taken in August 2026, scanned 79,848 listings and found 87.8% of agent skills have never had a security audit.
If you are weighing OmniRoute against the incumbent, we also published a comparison against LiteLLM (57,827 stars, MIT core plus a commercial enterprise tier, with its own MCP server listing): OmniRoute vs LiteLLM. Short version: OmniRoute belongs on a machine you own, LiteLLM on a proxy a team shares.
Where OmniRoute sits in the catalog
The repo ships 46 skills; skills.sh shows its 20 most-installed at 308 to 388 installs each (all-time, September 2, 2026), tiny numbers next to the head of the skills leaderboard where grill-me alone crossed a million. The 60k stars and the Breakout search curve say the gateway is what people want; the skills are along for the ride. That gap between repo fame and skill adoption is the same shape our agent economy census measured across the whole ecosystem: attention concentrates, installs concentrate harder.
Both entry-point skills are now listed on Skillselion, alongside the rest of the deep-dive series where we read one skill's actual file per piece.
Source repo: diegosouzapw/OmniRoute. The pinned file we quoted: skills/omni-auth/SKILL.md at 93265ee.
Disclosure: we run Skillselion, the directory these listings live on. The catalog is refreshed daily from skills.sh, GitHub and MCP registries, ranked by real installs.
Top comments (0)