LLM Wiki — Synaptic Lattice

Manifest (mnfst/manifest)

Properties1
tagsentity, repository, tool, llm-wiki, agentic, infrastructure

Leitura de 4 min

Overview

Manifest (manifest.build, repo mnfst/manifest) is an open-source LLM gateway — a proxy layer that sits between agents/harnesses and model providers, handling provider connections, request logging, observability, and error repair. 7.5k+ GitHub stars, TypeScript, MIT-licensed, self-hostable. The product’s identity shifted twice in 2026: it launched with an LLM router, deprecated the router in June 2026 (shut down September 1st), and pivoted toward Autofix — automatic repair of failing LLM/API requests — extending in September 2026 from LLM calls to APIs in general (“the self-healing layer for APIs”).

The router deprecation and its data-backed critique of on-the-fly model selection is documented in model-routing-llm.

Key Facts

  • What it is: open-source LLM gateway connecting agents and harnesses to any provider; all providers, all settings, no black box between agents and providers
  • Stars/license: 7,507 stars, TypeScript, MIT (GitHub API, 2026-09-04); repo created 2022-09-27
  • Autofix (July 2026): when a provider rejects a request with a fixable error (renamed parameter, deprecated model, unsupported field), the gateway applies a known-correction patch and retries — “repairs broken provider calls before your agent ever sees them”. Dashboard tracks requests run/failed/recovered, per-agent and global, with per-attempt detail showing the exact change applied
  • Router deprecated (June 2026, shutdown 2026-09-01): after four months across ~7,000 cloud users, results were mixed; see the anti-router position below
  • New direction (September 2026): pivoting from LLM-only to general API autofix — “Manifest is becoming the self-healing layer for APIs”
  • Deployment: Manifest Cloud (free plan + paid plans since July 2026) or self-hosted
  • Ecosystem: n8n community node (September 2026)

Anti-Router Position (2026-07-31)

The deprecation post is a rare named critique of the LLM-router trend — most rivals were launching routers while Manifest removed theirs. Bruno Perez’s three arguments, from 4 months of usage data (7,000 cloud users) rather than theory:

  1. Complexity cannot be deduced from the prompt alone — the prompt is just the trigger; complexity is discovered later via tool calls and web searches. “Evaluate the tests for the repo $GIT_REPO and improve them” ranges from trivial (personal HTML5 site) to massive (Linux kernel).
  2. Cache beats routing for cost reduction — cache reads are 75–90% cheaper than uncached inputs; system prompts and history sit at the prompt start where prefix caching works best. A cache-aware router ends up sticky to one model, “doing its job by, ironically, not doing it.”
  3. Routers break behavior consistency — engineers should choose models like craftsmen choose tools (“just as a painter knows exactly what brush”); jumping between models lowers work quality and fragments evals, system prompts, and observability into harder-to-maintain unpredictability. Per-request isolation with explicit model/params is “naturally superior in most cases”.

Implications

This is a vendor claim from a company that exited the router business, so self-interest discount applies — but it is data-backed (4 months, 7k users) and aligns with the core failure mode documented in model-routing-llm: task complexity only becomes knowable mid-execution. For this vault, the practical stance is explicit per-call model choice with caching, over on-the-fly complexity classification.

Relationships

  • “compatible-with” model-context-protocol — occupies the adjacent gateway/proxy layer: MCP standardizes tool access, gateways standardize provider access
  • “manages” agent-memory-systems — gateway observability (request logs, recovery stats) is the infrastructure side of agent reliability work
  • “differs-from” rag — unrelated functionally, but the same “cost per token” economics drive both cache-based and retrieval-based savings
  • “operates-with” claude-code — target harness type: any agent harness connecting to multiple providers

Gateway/observability infrastructure is the reliability layer this vault’s stack currently lacks — relevant if the owner’s multi-model usage (per resources) grows beyond one provider. The anti-router position is directly useful: it recommends explicit model selection per request, which matches how this vault already operates (deliberate model choice per task, documented in rank queries).

Sources

^[raw/external/manifest-build-5de6b15f.md] ^[raw/external/manifest-build-why-we-deprecated-our-llm-router-9e78a424.md] ^[raw/external/manifest-build-auto-fix-release-20f68c0f.md]

Visão de gráfico

Backlinks