sportsclaw in Machina

sportsclaw is the open-source side of Machina. Machina Sports builds and maintains it, and the platform runs the same engine inside every project as the sportsclaw relay. This page covers how the two fit together: what runs where, who calls what, and what is live versus still rolling out.

Callers

• Factory app builder
• Broadcast & studio desks
• Project connectors & workflows
• Your own services

authenticated HTTP · /api/query · /api/capabilities · /api/decide

Machina projectisolated tenant · provisioned per project

sportsclaw relayThe same open-source engine

• sportsclaw engine (TypeScript)
• sports-skills data layer · 17 default skills
• memory volume · per-user agent memory

MCP · SSE
project token

project MCP serverEverything the project owns

• licensed feeds via connectors
• documents, workflows, agents
• durable loop, when loop-runner is installed

alongside the project's Client API, workers, Redis and vault

Keyless public data

ESPN · FastF1 · Kalshi · Polymarket · news

Licensed & project data

real-time feeds, odds, the project's own documents

Outside-in works too: any sportsclaw install can join the same project with sportsclaw machina connect, which mints a durable service key through machina-cli.

Two ways in

Inside a project: the relay

When a project is created on Machina, the platform provisions it in an isolated tenant: an API, background workers, an MCP server, Redis and a vault — and a sportsclaw relay. The relay is this repository's engine running in relay mode (docker/relay), and it comes with:

• the sports-skills data layer, with 17 default skills covering sports, prediction markets, betting math, news and team metadata;
• a persistent memory volume;
• the project's MCP server, connected over SSE and authenticated with the project's own token;
• an HTTP API for Machina products and for your own services.

There is nothing to install or configure. The relay ships with the project.

From anywhere: sportsclaw machina connect

Any sportsclaw install — your laptop, a Discord bot, a container — can join a Machina project as a client:

pip install machina-cli && machina login
sportsclaw machina connect <project>

machina-cli resolves the project's MCP endpoint and mints a durable service key for sportsclaw (--org picks the organization, --probe tests the endpoint first). The key is stored in ~/.sportsclaw/.env, never in the config file. From then on your agent reads the project's tools next to the keyless skills. Details and manual setup: Licensed data & machina-cli.

Standalone first?Install sportsclaw and run your first grounded query in about a minute.Need a project?Licensed data, SLAs and the platform around the relay live at machina.gg.

What the relay serves

The relay is a small HTTP service in front of the engine. Each request runs the engine in an isolated child process with bounded time, output and concurrency.

Endpoint	What it does	Access
GET /health	Liveness	public
GET /api/skills	Installed skill catalog	public
POST /api/query · POST /api/query/sync	Run a question through the engine — streaming or blocking	X-Auth-Token
GET /api/capabilities	Protocol and engine version, build, skills, MCP allowlists, limits	X-Auth-Token
/api/agents	Create, list, update and delegate to native agents	X-Auth-Token
POST /api/decide	Typed Jev decisions, disabled by default	X-Auth-Token
/api/highlights/jobs	Highlight clipping jobs — preview	separate token

Authenticated routes fail closed: no token configured on the relay means 503, a missing or wrong token means 401. Prompt size, timeouts and concurrency are bounded per relay — see the relay contract.

Relays follow releases

A project's relay runs the image it was provisioned with, and relay features ship with relay-v* releases. Ask GET /api/capabilities for the engine version, skills and limits a relay actually runs before you depend on a newer endpoint.

Who calls it

• Factory — Machina's app builder gives its coding agent a sportsclaw_query tool, so it can pull live sports data through the relay while it writes your app.
• Broadcast and studio desks — desk copilots send questions to the relay and read /api/capabilities to show what it can do.
• Project connectors and workflows — a template connector can call the relay over HTTP like any other service, which is how project workflows reach it.
• Your own services — anything that can send an authenticated HTTP request.

The relay is a service you call. Being in the same project does not, by itself, give every platform agent sportsclaw's tools — callers opt in by calling the API.

Memory on the platform

sportsclaw keeps durable memory per user — SOUL, fan profile, context, reflections, strategy and conversation logs. With a Machina server connected, the default (auto) stores it as project documents through the pod driver, isolated per user and, when native agents are selected, per agent. Without one it falls back to local files; Hindsight is an explicit opt-in.

Callers that keep their own transcript can send history_mode: "caller". The relay then neither restores nor appends its own thread for that run, while durable memory stays scoped to the user_id.

Licensed data and the upgrade signal

The keyless skills read public APIs and are meant for personal, non-commercial use. Inside a project, licensed real-time feeds arrive as the project's own MCP tools, through connectors installed from Machina templates — the agent reads both side by side.

When a public source rate-limits or lacks a feed, the data layer attaches an upgrade hint. The agent mentions the licensed path at most once per conversation, never inside automated alerts or broadcasts, and never instead of an answer.

Durable work: the loop-runner

Some work outlives a single exchange. When a project has the durable loop installed — the loop-runner agent, provisioned with machina-cli — sportsclaw exposes a machina_loop tool with three actions: start, continue and read.

• Every session is a document on the project, resumed by a beat, so it survives restarts and async waits.
• A turn only closes as idle after a code check and an independent evaluator; otherwise it is marked for review.
• Starting or continuing a session asks for approval in sportsclaw, and replies come back as untrusted data.

See Durable Task Delegation.

Status at a glance

Capability	Status	Notes
Relay provisioned with every project	Live	automatic, no setup
Keyless sports and market skills	Live	17 by default
Project MCP tools inside the agent	Live	over SSE, with the project token
Pod memory	Live	default when a Machina server is connected
Authenticated query API and /api/capabilities	Live	from relay-v0.29.6
Durable loop delegation	Rolling out	needs loop-runner on the project
Typed decisions (/api/decide)	Opt-in	disabled by default
Highlight jobs	Preview	rights-gated, not generally available

Is sportsclaw the Machina platform?

No. sportsclaw is the open-source agent engine. Machina is the platform around it: isolated tenants, licensed data, templates, workflows, Studio, Factory, deploys and SLAs. sportsclaw runs perfectly well without Machina, and Machina runs much more than sportsclaw. They share a team, a data layer (sports-skills) and the engine you can read on GitHub.