Ryuksaidso is an agent control plane — a single place to define AI agents, execute them as durable jobs, pause their risky side effects for a human decision, and keep the evidence that proves they behave. It is the tooling you wish you had the first time an agent did something “weird” in production and nobody could say why.
What you get
Every run persists its planner, tool and synthesis steps with inputs, outputs, latency and token usage. Nothing lives only in a log line.
Policies pause dangerous tool calls until a person approves or rejects. Approvals are scoped, timed and written to the audit trail.
Draft freely, publish immutable versions with changelogs, pin one to production and roll back without redeploying anything.
Regression datasets score intent accuracy for every draft, so a version ships because it beat the last one — not because it felt right.
Postgres full-text retrieval keeps answers anchored in your own documents, scoped to your workspace.
Route between Ollama (local) and OmniRoute (cloud), switch models from the UI, and let the worker fail over automatically.
How the platform fits together
| Piece | Technology | Job |
|---|---|---|
| Web | Next.js 15 · React 19 | Control plane UI at :3000 |
| API | Express · Zod · Prisma | REST endpoints, auth, rate limits at :4001 |
| Worker | BullMQ on Redis | Executes runs, tools, approvals and evals |
| Database | PostgreSQL (pgvector) | Agents, runs, traces, knowledge, audit |
| Cache & queue | Redis 7 | Sessions, rate limits, durable job queue |
| LLM providers | Ollama · OmniRoute | Model inference for agents and evals |
Vocabulary
| Term | Meaning |
|---|---|
| Workspace | The tenant boundary. Every row in the database is scoped to one. |
| Project | A grouping for related agents, with its own production version pin. |
| Agent | Instructions, a tool allowlist and a knowledge scope. Editable draft. |
| Version | An immutable snapshot of an agent, published with a changelog. |
| Run | One execution of an agent version, queued as a durable job. |
| Step | A single action inside a run — plan, tool call or synthesis. |
| Policy | A rule that decides whether an action needs human approval. |
| Approval | The pending decision a human must make before a run continues. |
| Evaluation | A dataset run against an agent to score its behaviour. |