convez.ai
MCP ServerAutomaçãoAvançado
olliefms

Ollie — Sistema de Gestão de Frotas com MCP

olliefms/ollie

TMS self-hosted em Rust com MCP para agentes de IA e processamento documental

Ver no GitHub
InicialPluginScore · de 0 a 100

Sistema de Gestão de Transportes (TMS) self-hosted em Rust que gerencia cargas, viagens, motoristas e frotas com um servidor MCP integrado para agentes de IA. Processa documentos (BOLs, recibos, fotos) com OCR, embeddings semânticos e busca vetorial via Ollama e LanceDB.

Instalação

$npx convez-ai olliefms/ollie

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
1
Forks
0
Acompanhando
0
Saúde
75%
Último commit
há 2 meses
Versão
v2.10.0
Issues abertas
29
Contribuidores
3

Casos de uso

  • Automação de despacho de frotas com agentes de IA
  • Gerenciamento de cargas e viagens com rastreamento em tempo real
  • Busca semântica e Q&A sobre documentos operacionais (BOLs, recibos, PODs)
  • Integração de agentes Claude com operações de transporte

Veredito

Solução robusta e especializada para operações de transporte que expõe seu domínio via MCP, REST e portais dedicados. Exige infraestrutura Rust/Ollama, mas oferece controle total sobre dados sensíveis de frota em ambiente self-hosted.

README

A tradução para o português deste README ainda está na fila.

ollie

A self-hosted freight Transportation Management System (TMS) written in Rust.

ollie runs the operational core of a trucking dispatch operation — loads, trips, drivers, trucks, trailers, and facilities — and pairs it with an AI-enabled document store. It exposes three API surfaces (an MCP server for AI agents, a Fleet REST API, and a driver mobile portal) plus two bundled web apps: a Fleet SPA and a driver PWA.

Uploaded documents (rate cons, BOLs, PODs, photos) are stored content-addressed on disk and processed in the background by Ollama — text summaries and vector embeddings for semantic search, with images and scanned PDFs routed through a vision model. Vector search is powered by LanceDB.

What's in the box

DomainDescription
LoadsFreight load lifecycle: planned → assigned → dispatched → in_transit → delivered → invoiced → settled, with stops, rate line items, and detention tracking.
TripsDriver-facing execution of loads, with stop arrive/depart, check-calls, relay chaining, deadhead/loaded mileage (ORS HGV routing), and a full state machine.
Drivers / Trucks / TrailersFleet entities with status state machines and equipment attach/detach.
FacilitiesShippers/receivers, geocoded (US Census) with semantic dedup and search.
Documents (blobs)Content-addressed, deduplicated file storage with Ollama summaries, embeddings, and natural-language Q&A over a document.
EventsAppend-only journal of operational events.
DoctorsDiagnose-and-repair tools for trip / load / facility data integrity.

API surfaces

ollie exposes the same domain through three surfaces — pick by caller:

SurfacePathAuthUse
Fleet MCPPOST /fleet/mcpFleet user JWT or API keyAI agents and tool-using assistants. Preferred.
Fleet REST/fleet/api/v1/*Fleet user JWT or API keyFleet web app and programmatic integrations.
Driver portal/driver/api/v1/*Driver JWT (passkey or PIN)The driver mobile PWA only.

Three endpoints are public (no auth): GET /version, GET /openapi.json, and GET /llms.txt. Start with GET /llms.txt — it's a hand-written, agent-oriented tour of every surface, tool, and the domain model.

The web apps are served as static SPAs: the fleet app at /fleet and the driver PWA at /driver.

Prerequisites

  • Rust 1.78+ (for building from source)
  • Ollama running locally (for document AI features)
  • Docker + Docker Compose (for the containerized setup)
  • An OpenRouteService API key (for trip/load mileage)

Ollama models

Pull the document-processing models before starting:

ollama pull nomic-embed-text   # embeddings
ollama pull llama3.2           # text summaries
ollama pull moondream          # vision (images, scanned PDFs)

Quick start with Docker Compose

git clone https://github.com/olliefms/ollie.git
cd ollie

# Copy the example env and fill in the required secrets
cp .env.example .env
$EDITOR .env

docker compose up

The server listens on http://localhost:3000. The Compose file also starts an Ollama container and pulls the required models on first boot.

Building from source

git clone https://github.com/olliefms/ollie.git
cd ollie

cp .env.example .env
$EDITOR .env

cargo build --release
./target/release/ollie

Configuration

All configuration is via environment variables (or a .env file).

Required

VariableDescription
DRIVER_JWT_SECRETSigning secret for driver-portal JWTs. Min 32 bytes.
FLEET_JWT_SECRETSigning secret for fleet user JWTs and API keys. Min 32 bytes.
DRIVER_RP_IDWebAuthn relying-party ID for driver passkeys (e.g. localhost).
DRIVER_RP_ORIGINWebAuthn relying-party origin (e.g. http://localhost:3000).

Recommended

VariableDefaultDescription
ORS_API_KEY—OpenRouteService key. Without it, trip/load mileage is not computed.
OLLIE_PUBLIC_BASE_URLemptyExternally-reachable base URL (no trailing slash). Required for presigned blob upload/download URLs handed to MCP agents.

Optional

VariableDefaultDescription
PORT3000HTTP listen port
DATA_DIR./dataStorage root (blobs, extracts, LanceDB)
BLOB_STORE_PATH./data/blobsOverride blob storage path
EXTRACT_STORE_PATH./data/extractsOverride extracted-text cache path
LANCEDB_PATH./data/lancedbOverride LanceDB path
OLLAMA_BASE_URLhttp://localhost:11434Ollama API base URL
OLLAMA_EMBED_MODELnomic-embed-textEmbedding model
OLLAMA_SUMMARY_MODELllama3.2Text summarization model
OLLAMA_VISION_MODELmoondreamVision model (images, scanned PDFs)
OLLAMA_EMBED_DIM768Embedding dimension (must match the embed model)
PIPELINE_WORKERS1Concurrent document-processing workers
GEOCODING_WORKERS1Concurrent facility-geocoding workers
TERMINAL_TIMEZONEAmerica/New_YorkIANA timezone for driver pay-period weeks
OLLIE_FREE_DWELL_MINUTES120Free dwell before detention accrues at a stop
OLLIE_BLOB_PRESIGN_TTL_SECS300Default TTL for presigned blob URLs
OLLIE_BLOB_PRESIGN_MAX_TTL_SECS3600Hard cap on presigned blob URL TTL
FACILITY_DEDUP_HIGH_THRESHOLD0.92Cosine score above which a facility is an exact match
FACILITY_DEDUP_LOW_THRESHOLD0.75Cosine score above which a facility is a candidate match

Using the API

The fastest way to explore is the machine-readable spec and the LLM tour:

# Public — no auth required
curl http://localhost:3000/llms.txt         # agent-oriented overview of every surface
curl http://localhost:3000/openapi.json      # full OpenAPI 3.0 spec
curl http://localhost:3000/version           # { "version": "x.y.z" }

Fleet auth

# Log in with email + password to get a JWT
curl -X POST http://localhost:3000/fleet/auth/login \
  -H "Content-Type: application/json" \
  -d '{"email": "you@example.com", "password": "..."}'

# Use the returned token against MCP or REST
curl -X POST http://localhost:3000/fleet/mcp \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Headless callers can mint long-lived fleet user API keys under /fleet/api-keys (used in the Authorization header just like a JWT). See /llms.txt for the full MCP tool list, REST routes, auth model, and domain semantics (load/trip lifecycles, stop times and detention, facility resolution).

Document processing

Uploaded documents move through these states:

StatusMeaning
pendingQueued for AI processing
processingActively being summarized/embedded
readySummary and embedding complete
failedAI processing failed (check the error field)
File typeProcessing
Text, JSON, XMLExtracted and summarized as text
PDF (≥50 words)Text extracted, summarized as text
PDF (<50 words)Rendered as image, described by the vision model
ImagesDescribed by the vision model
OtherStored but not AI-processed

Files are content-addressed (SHA-256) and deduplicated — identical bytes share storage and AI output.

Development

cargo test       # unit + integration tests
cargo clippy     # lint
cargo build      # build

The integration tests do not require Ollama. See AGENTS.md for developer and AI-agent guidance, and docs/DESIGN.md for the UI design system.

Ver o original no GitHub

Bom para

Relacionados