A skill is a small folder — a prompt, an input/output schema, optional Python tools, and evals — that an agent runs end to end. This is the open-source runner that executes one entirely on your laptop: no Postgres, no bucket, no platform API, no sign-up. It's the same agent loop the hosted platform runs (one loop, two environments), so a skill behaves identically locally and in prod.

pip install "puras[local]"
puras run --local greeter --dir ./examples/hello-world -i name=Ada

• 🧑‍💻 Local-first — run and iterate on a skill before deploying anything.
• 🔌 Local API — puras serve exposes the hosted job API on localhost, so you build and test your app offline before you deploy.
• 🔑 Bring your own key — your provider, your bill, nothing billed by a platform.
• 🪶 Dependency-light — the offline path needs no DB/bucket/openai stack.
• 🔁 Prod parity — the same loop and contracts as Puras Cloud.

Getting started

Run it locally

The whole point of this repo — a skill's agent loop on your machine, on your key:

pip install "puras[local]"        # the puras CLI + the offline runner

# the bundled "hello world" skillpack has two skills: greeter + formatter
puras run --local greeter --dir ./examples/hello-world -i name="the Puras team"

It's BYO key: you call the provider directly and pay your own bill. The CLI reads your key from $ANTHROPIC_API_KEY (or the --api-key flag) and, if neither is set, prompts for it in the terminal.

From a checkout of this repo instead of PyPI:

pip install -e .             # puras-runner (the runtime)
pip install -e worker/sdk    # the puras CLI + SDK

…or use Puras Cloud (recommended for production)

The fastest way to run a skill with the full tool surface — media generation, web search/fetch, shared memory, persistent storage, durable resume, and eval suites at scale — is to sign up free at puras.co. No infrastructure to manage, automatic scaling, and a one-line MCP connect for Claude Code. The local runner here gives you the free, offline core; Cloud adds the managed, hosted surface for when you ship. The comparison below spells out exactly which is which.

Run a skill

puras run --local <skill> --dir <skillpack> -i KEY=VALUE [-i KEY2=VALUE2 ...]

• <skill> — the skill to run; omit it when the bundle has exactly one.
• --dir — the skillpack bundle root (a folder of <skill>/skill.yaml). Defaults to ..
• -i KEY=VALUE — an input, repeatable; validated against the skill's input_schema.
• --model claude/sonnet-4-6 — override the skill's model for this run.
• --api-key sk-... — your LLM key, if it isn't already in the environment.

Events stream to your console as the agent works; the final JSON output and a token tally (informational — you paid your provider, not Puras) print at the end.

Programmatic use is the same loop:

from worker.local_run import run_local

res = run_local("./examples/hello-world", {"name": "Ada"}, skill="greeter")
print(res["output"])

Run a skill's evals

Evals are to a skill what unit tests are to code. If a skill declares an evals: block, run its suite locally and gate on it:

puras eval --local content-repurposer --dir ./examples/content-studio --threshold 80

check / exact_match / schema graders run free; a rubric (LLM-as-judge) grader runs on your BYO key. --threshold N is a CI gate — non-zero exit if the pass-rate is below N.

Build your app against a local API

puras run --local answers "does my skill work?". When you're building the app that calls the skill, you want the other half: a local server that speaks the same API your app will hit in production. That's puras serve:

puras serve --dir ./examples/hello-world          # → http://127.0.0.1:8787

It mirrors the hosted job API (POST /v1/jobs, GET /v1/jobs/{id}, …/events, …/spans) backed by the offline runner — in-memory, zero extra dependencies. Point any Puras SDK at it by changing one thing — the base URL — and your app runs unchanged, offline, on your own key:

import puras

# api_base is the only thing that differs between local and prod
client = puras.Client(api_key="local", api_base="http://127.0.0.1:8787", skillpack="local")
print(client.run("greeter", {"name": "Ada"}))

import { Puras } from "puras";
const puras = new Puras({ apiKey: "local", apiBase: "http://127.0.0.1:8787", skillpack: "local" });
console.log(await puras.run("greeter", { name: "Ada" }));

…or just curl it:

curl -s "http://127.0.0.1:8787/v1/jobs?wait=true" \
  -H "content-type: application/json" \
  -d '{"skill": "greeter", "inputs": {"name": "Ada"}}'

When you ship, change the base URL to https://api.puras.co and puras deploy — the same app code now runs against the managed platform. Auth is open locally; --require-key <token> emulates API-key auth, and --host / --port change where it binds. (The Python and React-Native SDKs poll, so they work as-is; live SSE streaming is a Cloud feature.)

Build your own skill

cp -r examples/skillpack-template my-skillpack
$EDITOR my-skillpack/my-skill/SKILL.md      # the prompt
$EDITOR my-skillpack/my-skill/skill.yaml    # schema, model, tools, evals
puras run --local --dir ./my-skillpack -i topic=otters

my-skillpack/
  my-skill/
    SKILL.md          # the agent's instructions (system prompt)
    skill.yaml        # input/output schema + model + tools + evals
    tools/...         # optional deterministic Python tools the agent can call

When you're happy, the same bundle deploys to Puras Cloud unchanged. See the docs to deploy and call skills over the API.

Open-source vs Cloud

Puras is open-core: the runner — the agent loop and the local tool surface — is MIT-licensed and runs fully offline, forever. The hosted platform at puras.co is how the project is sustainably funded, and it adds the managed surface that can't exist on a single laptop. Premium isn't a crippled core — it's the capabilities that need real infrastructure.

The hosted-only tools (worker/agent_runner.py:PLATFORM_ONLY_TOOLS) are simply not offered to the model offline, so a skill that needs them still runs — it just won't see those tools locally. The exceptions are the three that have a BYO-key direct path (see below): the media generate_* verbs and web_search / web_fetch. The included examples (hello-world, skillpack-template, content-studio) use only the local surface and run end-to-end offline.

Media and web work offline too (BYO keys). The hosted platform owns a Fal key, a bucket, and a web service; the local runner instead reaches the same providers on your keys, with no platform in between. Set FAL_KEY and the generate_image / generate_video / generate_audio / transcribe verbs call Fal directly (worker/media_local.py) — same verb resolution and result shape as hosted, the output landing on your local drive (no bucket, no billing). And web_fetch runs as a plain HTTP GET while web_search runs through Anthropic's server-side web-search tool on your ANTHROPIC_API_KEY (worker/web_local.py), so a skill that researches the web works the same locally as in prod. Hosted adds managed billing, persistence to a bucket-backed drive, image search, and headless browser screenshots on top.

Workspace memory works offline too. The memory_search / memory_get / memory_put / memory_forget tools and the job-start memory injection are backed by a local SQLite file (worker/memory_store_sqlite.py) — the same agent-facing contract and the same hybrid ranking (exact identity + lexical, RRF-fused and shaped by recency × importance) as the hosted Postgres store, just without the pgvector semantic arm. So a skill's "shared brain" persists across puras run --local invocations. The DB lives next to the local drive root by default; point it anywhere with PURAS_LOCAL_MEMORY_PATH. Hosted memory adds semantic (embedding) retrieval and cross-workspace scope on top.

How it works

The runner runs the agent on a LocalRunContext with platform_enabled=False (but memory_enabled=True — its memory_backend() returns the SQLite store). That RunContext seam is the whole trick: one agent loop, two environments. The offline import path is kept dependency-light — no Postgres/bucket/openai at import time — and that's enforced by tests/dry/test_local_import_isolation.py.

Community & contributing

Questions, bugs, and skill ideas are welcome in Issues and Discussions. PRs that improve the runner, the docs, or the examples are appreciated.

License

MIT. The hosted platform's server-side code is separate and commercial — this runner, the SDK, and the examples are MIT and yours to use, modify, and self-run.