Ryde

Crash-proof your AI agent runs. Never re-pay for LLM calls you already made.

ryde resume picks up exactly where your agent crashed — skipping every step it already completed, and telling you exactly how much money that saved.

Install

```bash npm install -g ryde-cli

or run directly without installing:

npx ryde-cli submit your-workflow.ts ```

The problem

Agent loops call LLMs 5-10+ times per run. If the process crashes on step 4, re-running means re-paying and re-waiting for steps 1-3 you already finished.

30-second demo (no API key needed)

```bash git clone https://github.com//ryde.git cd ryde npm install docker run -d --name ryde-pg -e POSTGRES_PASSWORD=pass -p 5432:5432 postgres:16 docker exec -it ryde-pg psql -U postgres -c "CREATE EXTENSION IF NOT EXISTS pgcrypto;"

npm run ryde submit examples/agent-job.ts

Ctrl+C after a couple steps complete

npm run ryde resume examples/agent-job.ts

already-completed steps skip instantly — no re-spend

```

How it works

Every ctx.step(name, fn) call is checkpointed to Postgres with its result. On resume, the workflow re-runs from the top, but any step already marked completed short-circuits and returns its cached output. Same core pattern Temporal/Cadence use internally, scoped to single-node + Postgres so it runs in one Docker container with zero infra setup.

Bring your own LLM provider

Ryde doesn't call any LLM itself and needs no API key to install or run — it's a generic step-checkpointing engine. To use it with a real agent:

1. Write a workflow file, wrap each LLM/tool call in ctx.step(name, fn)
2. Inside fn, call whatever provider you want — OpenAI, Claude, local models
3. Set your own API key as an environment variable — Ryde never sees or needs it

See examples/agent-job-openai.ts and examples/agent-job-claude.ts for ready-to-adapt templates.

Roadmap

• [ ] Concurrent workers + row-level locking
• [ ] Web dashboard: live execution timelines, cost saved per resume
• [ ] Leader election for multi-worker coordination