Self-hosted general agent — every decision constrained, auditable, replayable.

Reyn is a self-hosted agent runtime whose entire decision loop is an OS-enforced contract — so every run is predictable, auditable, and replayable. It speaks the standard agent protocols (MCP, A2A), but bets on the loop itself being yours to verify rather than on integration breadth.

git clone https://github.com/tya5/reyn.git
cd reyn && pip install -e ".[dev]"
reyn init
reyn chat          # talk to the agent — persistent memory, RAG recall, MCP + A2A built in

---

What it guarantees

Most agent frameworks optimize for reach. Reyn optimizes for the integrity of the loop itself — what the runtime promises about every decision the model makes.

Guarantee	What it means
Bounded decisions	The model only picks from options the OS hands it (next step + a typed result). It can't invent a step, jump somewhere unlisted, or skip validation — a bad or hallucinated output is rejected before any side effect happens.
Everything on disk, replayable	Data lives in one workspace; every state change is appended to a log. Crash recovery and full run-replay come from those files — never from in-memory state.
Spending you can cap	Token and dollar limits per agent / chain / model. A run refuses to continue before it overspends, so there are no surprise bills from runaway loops.
Isolated & auditable	Each automation run sees only its own credentials (it can't reach another run's secrets), and every logged event carries an agent identity — ready for SOC 2 / ISO 27001 / audit trails.
Weak models stay viable	Because the OS absorbs capability gaps structurally, low-cost models run agent workflows reliably without prompt-level workarounds.

The trade-off is explicit: predictability and auditability over maximum autonomy. If you want the densest integration ecosystem and maximum LLM latitude, a connectivity-first agent or a workflow framework will feel less restrictive.

These guarantees are formalized as the OS's eight design principles — see principles.

---

See it

_{reyn chat — ask, get a streamed reply. (Animated walkthrough via vhs coming.)}

Cost & token transparency (Ctrl+B)	Time-travel (/rewind)

---

Quick Start

Requirements: Python 3.11+, a LiteLLM-compatible model endpoint.

pip install -e .               # local install; web UI: pip install -e ".[web]"
export OPENAI_API_KEY=sk-...   # or the key for your LiteLLM proxy
reyn init                      # creates reyn.yaml + .reyn/config.yaml
reyn chat
> What can you do?

reyn chat opens a session with persistent per-agent history, tool-calling (file ops, web, MCP), and automatic recall over any indexed source.

A note on the default model. Reyn's default models.standard points at a low-cost LLM; occasional empty replies on tool-heavy queries and router-vocabulary leakage in non-English answers are normal at the weak tier and dissolve on stronger models. Point models.standard at a stronger model if it matters — details in the Quick Start guide.

Index your own docs for retrieval, or write a reusable typed workflow:

reyn run index_docs '{"type":"index_docs_input","data":{"source":"my_docs","path":"docs/**/*.md","description":"Project documentation"}}'
reyn run my_skill "Summarize AI trends in education."

Walkthroughs: chat mode · RAG.

---

Architecture

A request flows through a small set of durable parts; the OS sits in the middle and writes everything it does to three stores of record.

flowchart TB
    User([User]) --> Agent["Agent · carries agent_id"]
    Agent --> S
    subgraph S["Session: binds identity, capability and scope"]
      direction LR
      Cap["Capability<br/>automations · MCP · A2A · RAG · tools"]
      Perm["Permission<br/>per-run credential scope"]
    end
    S --> OS{{"OS runtime — bounded-decision loop"}}
    Cap -. invoked via .-> OS
    Perm -. gates .-> OS
    OS --> WS[("Workspace<br/>data · single source of truth")]
    OS --> WAL[("WAL<br/>crash recovery")]
    OS --> AE[("AuditEvent log<br/>append-only · replayable")]
    classDef store fill:#1f2937,color:#fff,stroke:none;
    class WS,WAL,AE store;

Loading

• Actors — the User talks to an Agent (which carries an audit identity), inside a Session that binds what the agent can do and is allowed to do.
• Capability — what the agent can do: automations, MCP/A2A peers, RAG recall, tools.
• Governance — Permission gates every action; cost caps and the sandbox bound it.
• State of record — the Workspace holds data, the WAL drives crash recovery, the AuditEvent log drives replay and audit. These are separate logs on purpose.

The diagram above is the depth view — one agent's execution. The breadth view shows how agents are reached and how they relate: every trigger converges on an agent's inbox, and the fleet delegates under a topology.

flowchart LR
    HUMAN["Human (direct)<br/>TUI · Web UI"]
    subgraph EXT["External connection — outside calls in"]
      direction TB
      MCP["MCP server<br/>AI clients"]
      A2A["A2A server<br/>peer agents"]
      GW["Gateway<br/>Slack / LINE webhooks"]
    end
    INT["Internal trigger<br/>cron"]
    INBOX(["agent inbox<br/>send_to_agent"])
    EXT --> INBOX
    INT --> INBOX
    subgraph FLEET["Agent fleet"]
      direction TB
      REG[("AgentRegistry<br/>multi-agent")]
      Ax["Agent A · agent_id"]
      Ay["Agent B"]
      Ax --- SAx["Sessions<br/>multi-session"]
      REG --> Ax
      REG --> Ay
    end
    HUMAN -- submit_user_text --> Ax
    INBOX --> FLEET
    HOOK["In-turn intervention hooks<br/>(proposed)"] -. block / rewrite running turn .-> Ax
    Ax <-. "delegate · topology-gated<br/>network / team / pipeline · chain_id" .-> Ay
    classDef pr stroke-dasharray:5 4;
    class HOOK pr;

Loading

• Reached two ways — humans drive a session directly (TUI / browser Web UI → submit_user_text), while systems and schedules inject into an agent's inbox: external connection (MCP server · A2A server · gateway webhooks like Slack / LINE) and internal trigger (cron). An in-turn hook layer can intercept a running turn (proposed). The HTTP surfaces — A2A, MCP-over-SSE, gateway webhooks, and the browser Web UI — are all served by the one reyn web (FastAPI) gateway. See interaction layers.
• Related — an AgentRegistry holds many Agents (each an identity), each with many parallel Sessions; agents delegate to peers under a topology (network / team / pipeline), hop-capped and chain_id-traced. See multi-agent · sessions.

Details: architecture · principles.

---

How it compares

Reyn overlaps with three families of agent tooling but makes a different bet — the loop itself as an OS-enforced contract, not connectivity reach or programmable flexibility. Each peer below is described by what it's great at, not by what it lacks.

General & coding agents — the closest peers:

Tool	Category	Primary bet — what it's great at
OpenClaw	Self-hosted personal assistant	Multi-channel reach — answers you on the channels you already use (WhatsApp / Telegram / Slack …), on your own devices
Hermes	Self-hosted AI agent	A built-in self-improving loop — creates and refines skills from experience, curates its own memory (Nous Research)
Claude Code	Terminal coding agent	Autonomous multi-file coding that lives in your terminal
Codex	Terminal coding agent (OpenAI)	A local coding agent with IDE / desktop / cloud variants
Reyn	Self-hosted agent OS	The loop as an OS-enforced contract — bounded decisions, a replayable audit log, and per-agent cost caps

Reyn connects too (MCP + A2A) and runs open-ended tasks, but optimizes for integrity of the loop rather than connectivity breadth or a turnkey coding workflow.

Workflow frameworks — if you'd rather wire the loop yourself:

Framework	Loop enforcement	State & replay	Strength
LangGraph	Code-defined graph; LLM can pick arbitrary transitions via Command()	Checkpointer (SQLite/Postgres); time-travel	Expressiveness; LangChain ecosystem
CrewAI	Role-driven; no OS-level candidate constraint	@persist; task replay (last run)	Role-orchestration ergonomics
AutoGen	Conversational; LLM selects next speaker freely	App-managed state; OTel spans (no replay)	Multi-agent dialog patterns
Semantic Kernel	Function-calling loop; autonomous plugin selection	In-memory history; OTel spans (no replay)	Azure-native; C#/Python/Java parity
Reyn	OS-enforced: validated transitions, closed candidate set	Workspace + WAL; append-only replayable events	Predictability; audit trail; weak-model viability; cost caps

Reyn fits when every LLM decision must be replayable and auditable, when you want weak models to be reliable, or when you need hard cost caps. Reyn doesn't fit quick max-flexibility prototyping, single-shot stateless calls, or a turnkey RAG/UI product — those live downstream (see care boundary).

---

Documentation

📖 Read the docs · 🏠 Project site — MkDocs Material, English + Japanese (in progress).

Section	What's covered
Getting started	Install, chat mode, evals
For users	Day-to-day usage: permissions, sandbox, cost caps, scheduling, memory
For automation authors	Phases, composition, MCP, validation, operations
Reference	CLI, config (reyn.yaml), Control IR, events
Concepts	Architecture, principles, the act-sense-react loop
Feature inventory	Every implemented feature, grouped by subsystem

make docs-install && make docs-serve   # http://127.0.0.1:8000

---

Integrations

Reyn both consumes external tools and exposes itself to other systems over standard protocols — no extra process or port; the same gateway backs the web UI.

• MCP — talk to a Reyn agent from any MCP client (Claude Desktop, Claude Code, Cursor, …) via reyn web (SSE) or reyn mcp serve (stdio). See reyn mcp reference.
• A2A — expose Reyn agents as addressable A2A peers for other agents to discover and converse with, including async tasks and mid-run ask_user. See A2A concepts.
• Self-reading — the agent can browse its own repository (reyn_src_list / reyn_src_read) to answer "how does Reyn work?" from source.

---

Project Status

1.0 OSS launch ready. The framework foundation is green and dogfood discipline is operational: stable DSL, CLI, and event-log surfaces; an authentication stack (MCP bearer headers, OAuth refresh, per-run credential scoping, agent-id propagation); and a RAG framework foundation.

Deliberate maturity gaps live downstream — vector stores beyond SQLite, advanced retrieval (rerank / HyDE), a RAG eval framework, IDE integration, and sensitive-data redaction. The "framework foundation" framing is honest, not a hedge; see Project Status in the docs and CLAUDE.md for the full list and architectural constraints.

---

License

MIT — see LICENSE.

Acknowledgements — powered by AI

Reyn is powered by AI both at runtime (every workflow execution delegates decisions to an LLM via LiteLLM) and in its development (substantial portions of the code, stdlib components, docs, and website were drafted with AI tooling — primarily Claude Code, with human review and the final architectural calls held by the maintainer). This disclosure is mandatory rather than promotional: for provenance, start with git log --grep="Co-Authored-By: Claude" and the design prompts under website/_design/.