README.md
Rama main · réplica de lectura del repositorio original.
<div align="center">
# OpenHermit
**Agents, but operable.**
Open-source platform for deploying fleets of AI agents as production services — durable state, sandboxed execution, managed at scale, and the channels you already use.
Built by [**Amiko**](https://heyamiko.com) — an agent-agnostic infrastructure platform for the AI era.
[](https://www.npmjs.com/package/openhermit)
[](LICENSE)
[](https://github.com/HCF-S/openhermit/stargazers)
[](https://www.typescriptlang.org/)
[](#contributing)
[**Website**](https://openhermit.ai) · [**Why OpenHermit**](https://openhermit.ai/blog/agents-but-operable) · [**Docs**](docs/) · [**npm**](https://www.npmjs.com/package/openhermit)
</div>
---
## Why OpenHermit
Most CLI-based agents (Claude Code, OpenClaw, Hermes, …) keep their state in files: memories as markdown, sessions as JSONL, skills as folders, secrets as dotfiles. That's perfect for one human at one machine — but it falls apart the moment you stop being one human. Running an internal agent platform for your team, a SaaS where every customer gets their own agent, or a swarm of specialized roles? Files scatter, secrets leak, fleet operations turn into SSH-and-pray.
OpenHermit makes one core design choice: **separate internal state from external state.**
- **Internal state** — sessions, memories, instructions, skills, MCP servers, schedules, secrets, users — lives in shared **PostgreSQL**, scoped by `agent_id`.
- **External state** — the workspace files an agent is currently working on — lives in a per-agent **sandbox**. Pick the backend that fits your deployment: a self-hosted Docker container, or a cloud sandbox provider like [E2B](https://e2b.dev) or [Daytona](https://www.daytona.io).
Once internal state is centralized, fleet operations become trivial:
```bash
hermit skills enable standup-digest --all # roll a skill out to every agent
hermit mcp enable mcp_github --all # add an MCP server to every agent
hermit instructions append rules "Never share PII." --all # push a rule to every agent
hermit config secrets set OPENROUTER_API_KEY sk-... --agent main # rotate a secret
```
📖 [Read the full reasoning →](https://openhermit.ai/blog/agents-but-operable)
---
## Features
- 🚪 **Gateway control plane** — single Hono server. Agents start, attach, detach without orchestration. Admin UI at `/admin/`.
- 🐘 **Postgres-backed state** — sessions, memories, instructions, skills, MCP, schedules, secrets — durable behind Drizzle.
- 🐳 **Sandboxed execution** — per-agent sandbox: self-hosted Docker, E2B, Daytona, or [Tenki](docs/sandbox-model.md#tenki). Code runs isolated from the gateway, with the same exec interface across backends.
- 💬 **Channels included** — Telegram, Discord, Slack adapters, package-installed Signal / WeChat / WhatsApp, plus CLI and Web UI. Enable, disable, reconfigure at runtime.
- 🛠 **Skills & MCP servers** — install centrally, enable per-agent or fleet-wide, audit from one place.
- ⏱ **Schedules & automation** — cron and one-shot jobs with timeout, concurrency policy, and error backoff.
- 👥 **Multi-user with roles** — owner / user / guest. Identity reconciliation across CLI, web, and channels.
- 🔌 **Multi-protocol transport** — HTTP sync, inline SSE streaming, durable SSE, WebSocket RPC.
- 🔬 **Deep Research** — durable, session-attached research runs: an LLM-drafted plan you review and approve, a bounded search/read/extract loop with hard budgets, and a final report whose every finding cites server-verified evidence excerpts. Pause, refine, resume; runs survive restarts. Requires a configured web provider (Tavily or Exa recommended; the default Defuddle provider works but scrapes Google results). Current limitations: HTML/text sources only (PDFs are reported as unsupported), web-only sources (no uploaded files or MCP yet), manual resume after an unclean restart. **Note: research runs are model-intensive — see [cost and model selection](#deep-research-cost-and-model-selection) before your first run.** Design: [Deep Research Design](docs/deep-research-design.md).
---
## Deep Research: cost and model selection
A research run is a pipeline of model calls — planner, per-iteration action
decisions, one evidence extraction per source read (these are the big ones:
each carries up to a 200 KB page snapshot ≈ 50k input tokens), and synthesis.
Budgets are hard caps per depth; typical runs land well under them, but plan
for the worst case:
| Depth | Model calls | Input tokens | Output tokens | e.g. gpt-5-mini | e.g. gemini-3-flash | e.g. claude-sonnet-4-5 |
|---|---:|---:|---:|---:|---:|---:|
| `quick` | ≤ 22 | ≤ 250k | ≤ 40k | ~$0.14 | ~$0.25 | ~$1.35 |
| `standard` | ≤ 40 | ≤ 500k | ≤ 80k | ~$0.29 | ~$0.49 | ~$2.70 |
| `thorough` | ≤ 72 | ≤ 900k | ≤ 120k | ~$0.47 | ~$0.81 | ~$4.50 |
Dollar figures are worst-case at list prices. Search costs are separate but
small (Tavily's free tier of ~1,000 credits/month easily covers testing).
**Free model tiers are usually not enough.** The binding constraint is
per-minute/per-day token quotas, not intelligence — a single extraction call
can be ~50k input tokens, which blows through e.g. Google AI Studio's free
tier mid-run (you'll see a 429 quota error surfaced on the run). Use a paid
key, or a topped-up OpenRouter account (~$5 covers many standard runs).
**What a model needs to handle research well:** disciplined JSON output
(every phase is validated, with one repair attempt), a context window of
64k+ tokens for extraction, and `model.max_tokens` of at least ~16k so long
extraction/synthesis answers don't truncate. Suggested configurations
(verified registry ids; prices per Mtok in/out):
| Provider / model | Price | Notes |
|---|---|---|
| `openai` / `gpt-5-mini` | $0.25 / $2 | Best price/reliability balance for research phases |
| `google` / `gemini-3-flash-preview` | $0.50 / $3 | Fast; needs a **paid-tier** key; keep `thinking low` |
| `anthropic` / `claude-haiku-4-5` | $1 / $5 | Very reliable JSON discipline |
| `anthropic` / `claude-sonnet-4-5` | $3 / $15 | Highest report quality; use for runs that matter |
| `openrouter` / `deepseek/deepseek-chat-v3.1` | $0.15 / $0.75 | Cheapest credible option; 32k context is tight for big pages |
Set with:
```bash
hermit config set model.provider openai
hermit config set model.model gpt-5-mini
hermit config set model.max_tokens 16384
hermit config secrets set OPENAI_API_KEY sk-...
```
### Local models (no subscription)
Any OpenAI-compatible local server works — Ollama, LM Studio, llama.cpp
`--server`, vLLM — via a custom endpoint:
```bash
# Ollama example. Research needs long context: raise it from Ollama's default.
OLLAMA_CONTEXT_LENGTH=65536 ollama serve # in another terminal: ollama pull qwen3:32b
hermit config set model.provider ollama
hermit config set model.model qwen3:32b
hermit config set model.base_url http://127.0.0.1:11434/v1
hermit config set model.api openai-completions
hermit config set model.max_tokens 16384
hermit config secrets set OLLAMA_API_KEY ollama # any non-empty value; local servers ignore it
```
Honest expectations for local runs:
- **Model size matters.** The phases demand strict JSON against non-trivial
schemas. 27B+ models (`qwen3:32b`, `gemma3:27b`, `llama3.3:70b`) are
workable; 7–8B models fail validation often enough to be frustrating even
with the built-in repair attempt and advisory-field tolerance.
- **Context is the hard requirement.** Extraction feeds up to ~50k tokens of
page snapshot; with a small context window the page gets cut and evidence
quality drops. Configure 64k+ (`OLLAMA_CONTEXT_LENGTH`, `--ctx-size`, …).
- **Time adds up.** A quick-depth run is ~15–25 sequential-ish calls, several
with huge prompts. On CPU-only or modest-GPU hardware expect a run to take
an hour or more where a hosted model takes minutes.
- Web search is unaffected — pair a local model with Tavily's free tier (or
the keyless default provider) for a fully subscription-free setup.
---
## Architecture
- **Admin** — CLI + Web UI for deploying and operating agents.
- **Client** — Web and CLI for end-users to chat with agents.
- **Channels** — Telegram, Discord, Slack, and package-installed adapters such as Signal, WeChat, and WhatsApp wired into any agent.
- **Gateway** — API, auth, routing, agent lifecycle, schedules.
- **Agent** — Model loop, tools / skills / MCP, sandboxed workspace (Docker / E2B / Daytona).
- **Storage** — PostgreSQL for every kind of internal state.
---
## Installation
```bash
npm install -g openhermit
```
This installs both `hermit` and `openhermit`.
For local development:
```bash
git clone https://github.com/HCF-S/openhermit.git
cd openhermit
npm install
```
---
## Quick Start
```bash
# Configure DATABASE_URL, GATEWAY_ADMIN_TOKEN, GATEWAY_JWT_SECRET.
hermit setup
# Start the gateway and the end-user web app.
hermit gateway start
hermit web start
# Check platform health.
hermit status
hermit doctor
# Create and start an agent.
hermit agents create main
hermit agents start main
# Chat through the CLI.
hermit chat --agent main
```
The gateway defaults to `http://127.0.0.1:4000` and serves the admin UI at `/admin/`. The end-user web app runs on `http://127.0.0.1:4310`.
---
## CLI Reference
| Area | Commands |
|------|----------|
| Setup | `hermit setup` |
| Gateway | `hermit gateway start`, `stop`, `run`, `status` |
| Web | `hermit web start`, `stop`, `run`, `status` |
| Agents | `hermit agents list`, `create`, `start`, `stop`, `restart`, `delete` |
| Chat | `hermit chat`, `--agent <id>`, `--resume`, `--session <sessionId>` |
| Config | `hermit config show`, `get`, `set` |
| Secrets | `hermit config secrets list`, `set`, `remove` |
| Instructions | `hermit instructions list`, `get`, `set`, `append`, `remove` — single-agent (`--agent <id>`) or admin fan-out (`--all`) |
| Skills | `hermit skills list`, `assignments`, `scan`, `register`, `delete`, `enable`, `disable` |
| MCP | `hermit mcp list`, `assignments`, `enable`, `disable` |
| Schedules | `hermit schedules list`, `create`, `pause`, `resume`, `delete`, `runs` |
| Operations | `hermit status`, `hermit stats`, `hermit doctor`, `hermit logs [-f] [-n N]` |
Agent-scoped commands accept `--agent <id>` and default to `OPENHERMIT_AGENT_ID` or `main`. Full reference: [docs/cli.md](docs/cli.md).
---
## API Overview
Agent execution routes are exposed under `/api/agents/{agentId}`:
- `POST /api/agents/{id}/sessions`
- `GET /api/agents/{id}/sessions`
- `POST /api/agents/{id}/sessions/{sessionId}/messages`
- `POST /api/agents/{id}/sessions/{sessionId}/messages?wait=true`
- `POST /api/agents/{id}/sessions/{sessionId}/messages?stream=true`
- `GET /api/agents/{id}/sessions/{sessionId}/events`
- `POST /api/agents/{id}/sessions/{sessionId}/approve`
- `POST /api/agents/{id}/sessions/{sessionId}/checkpoint`
- `DELETE /api/agents/{id}/sessions/{sessionId}`
- `ws://host/api/agents/{id}/ws`
Admin and owner-facing management endpoints live under `/api/admin/...` and `/api/agents/{agentId}/...`. Channel webhooks land at `POST /api/agents/{id}/channels/{namespace}/webhook`.
See [docs/transport-protocol.md](docs/transport-protocol.md), [docs/skills.md](docs/skills.md), [docs/mcp-servers.md](docs/mcp-servers.md), and [docs/channel-adapter.md](docs/channel-adapter.md).
---
## Internal State
All durable internal state is scoped by `agent_id` where applicable:
| Store | Contents |
|-------|----------|
| Agents | Registered agents, runtime config, security policy, workspace dirs |
| Sessions | Metadata, status, participants, working memory, descriptions |
| Session events | User, assistant, tool, error, channel, and introspection events |
| Memories | Long-term memory with PostgreSQL FTS plus ILIKE fallback |
| Instructions | Agent identity, behavior, and rules included in prompts |
| Users | Users, identities, roles, and merge links |
| Sandboxes | Per-agent sandbox rows (docker / e2b / daytona) with lifecycle and runtime state |
| Skills | Skill library and per-agent/global assignments |
| MCP servers | External MCP server definitions and assignments |
| Channels | Built-in and external channel rows with encrypted tokens |
| Channel credentials | Encrypted channel-owned auth state such as WhatsApp Web / Baileys credentials |
| Secrets | Per-agent provider/integration secrets, encrypted at rest |
| Schedules | Cron/once jobs and run history |
Secrets and channel-owned credentials are encrypted with `OPENHERMIT_SECRETS_KEY` (AES-256-GCM); without that key the gateway falls back to per-agent `secrets.json` for local dev secrets, but DB-backed channel credentials are unavailable. The only per-agent files on disk are the workspace at `~/.openhermit/workspaces/{agentId}/`; enabled skills are synced into each backend's own `<agent_home>/.openhermit/skills/system/` (bind-mounted for docker, uploaded via SDK for e2b/daytona).
---
## Repository Structure
```text
openhermit/
├── apps/
│ ├── agent/ # AgentRunner, tools, runtime, scheduler, channels
│ ├── gateway/ # Control plane, auth, admin API, admin UI
│ ├── cli/ # Published `hermit` / `openhermit` CLI
│ ├── web/ # End-user browser chat app
│ └── channels/
│ ├── telegram/ # Telegram adapter
│ ├── discord/ # Discord adapter
│ └── slack/ # Slack adapter
├── packages/
│ ├── protocol/ # Shared protocol types and route builders
│ ├── sdk/ # HTTP/SSE/WebSocket clients
│ ├── shared/ # Common env, errors, URL helpers
│ └── store/ # Drizzle schema and PostgreSQL store implementations
├── skills/ # Built-in OpenHermit skills registered by the gateway
└── docs/ # Architecture and operation docs
```
---
## Development
```bash
npm run dev:gateway # Gateway and admin UI API at http://127.0.0.1:4000
npm run dev:web # End-user web app at http://127.0.0.1:4310
npm run dev:cli # CLI from source
npm run dev:studio # Drizzle Studio for the configured database
npm run typecheck # Type-check all workspaces
npm test # Build and run test suites
```
Important environment variables:
| Variable | Description |
|----------|-------------|
| `DATABASE_URL` | PostgreSQL connection string |
| `DATABASE_URL_TEST` | Test PostgreSQL connection string used by `npm test` |
| `GATEWAY_ADMIN_TOKEN` | Bearer token for admin APIs and CLI management |
| `GATEWAY_JWT_SECRET` | JWT signing secret for user/device tokens |
| `GATEWAY_HOST` | Gateway listen host, default `127.0.0.1` |
| `GATEWAY_PORT` / `PORT` | Gateway port, default `4000` |
| `OPENHERMIT_SECRETS_KEY` | AES-256-GCM key used to encrypt `agent_secrets` and channel tokens at rest |
| `OPENHERMIT_TOKEN` | CLI token, usually the admin token |
| `OPENHERMIT_GATEWAY_URL` | Gateway URL, default `http://127.0.0.1:4000` |
| `OPENHERMIT_AGENT_ID` | Default CLI agent ID, default `main` |
| `OPENHERMIT_WEB_PORT` | End-user web app port, default `4310` |
| `AWS_ACCESS_KEY_ID` / `AWS_SECRET_ACCESS_KEY` / `AWS_REGION` | Standard AWS credential chain — used when `attachments.storage.provider = s3` |
| `SUPABASE_URL` | Supabase project URL (e.g. `https://xyz.supabase.co`) — required when `attachments.storage.provider = supabase`. Treated as part of the credential bundle since it embeds the project ID |
| `SUPABASE_SERVICE_ROLE_KEY` or `OPENHERMIT_ATTACHMENT_SUPABASE_SECRET_KEY` | Supabase secret API key (`sb_secret_…`) — required when `attachments.storage.provider = supabase`. Never stored in gateway config. The legacy service-role JWT still works. (`OPENHERMIT_ATTACHMENT_SUPABASE_SERVICE_KEY` is deprecated) |
### Attachment storage backends
The gateway supports three storage providers. The provider is selected in the gateway config (DB-backed, edited via the admin UI at `/admin/config` — JSON tab — or seeded from `gateway.json` on first boot), and **credentials live in env, not config** — only secrets go in env; non-secret pointers (provider, bucket, region, prefix, endpoint, root) live in the gateway config.
#### Local disk (default)
No config block needed; defaults to `~/.openhermit/attachments`. To override the root, set this under `attachments.storage`:
```json
{
"attachments": { "storage": { "provider": "local", "root": "/srv/openhermit/attachments" } }
}
```
#### S3 (or any S3-compatible service: R2, MinIO, B2)
Install the SDK on the gateway: `npm install @aws-sdk/client-s3 @aws-sdk/s3-request-presigner`. Credentials come from the AWS default chain (env vars / IAM role / IRSA).
```json
{
"attachments": {
"storage": {
"provider": "s3",
"bucket": "oh-attachments",
"region": "us-east-1",
"prefix": "prod",
"endpoint": "https://abc.r2.cloudflarestorage.com",
"forcePathStyle": false
}
}
}
```
#### Supabase Storage
Install the SDK on the gateway: `npm install @supabase/supabase-js`. Set `SUPABASE_URL` and a Supabase secret API key (`sb_secret_…`) in env as `SUPABASE_SERVICE_ROLE_KEY` (or `OPENHERMIT_ATTACHMENT_SUPABASE_SECRET_KEY`) — the project URL embeds the project ID and is treated as part of the credential bundle. The legacy service-role JWT still works; the older `OPENHERMIT_ATTACHMENT_SUPABASE_SERVICE_KEY` var is deprecated.
```json
{
"attachments": {
"storage": {
"provider": "supabase",
"bucket": "attachments",
"prefix": "agents"
}
}
}
```
Optional `attachments.limits.maxBytes` overrides the default 25 MB cap. Every successful upload is materialized into the sandbox; if the sandbox is down the row is marked `failed` and `attachment_fetch` re-materializes on first read.
---
## Documentation
- [CLI Reference](docs/cli.md)
- [Architecture](docs/architecture.md)
- [Plugins & Hooks (design draft)](docs/plugins.md)
- [Storage Model](docs/storage-model.md)
- [Session Model](docs/session-model.md)
- [User Model](docs/user-model.md)
- [Memory Model](docs/memory-model.md)
- [Sandbox Model](docs/sandbox-model.md)
- [Transport Protocol](docs/transport-protocol.md)
- [Access Policy](docs/access-policy.md)
- [Tools](docs/tools.md)
- [Skills](docs/skills.md)
- [MCP Servers](docs/mcp-servers.md)
- [Channel Adapters](docs/channel-adapter.md)
- [Introspection Design](docs/introspection-design.md)
- [Deep Research Design](docs/deep-research-design.md)
- [Architecture Decisions](docs/decisions.md)
- [Shipped Features](docs/plan.md)
- [Roadmap](docs/roadmap.md)
- [Open Questions](docs/pending-decisions.md)
---
## About Amiko
OpenHermit is the open-source agent infrastructure layer of [**Amiko**](https://heyamiko.com) — an agent-agnostic platform that integrates identity, productivity, social, and economic infrastructure into one coherent system where humans and AI agents coexist.
Amiko is organized around four connected layers:
- **Identity** — persistent, evolving user representation that powers personalization across all agents
- **Productivity** — agent capability integration across tools, services, and environments
- **Social** — 24/7 social continuity through AI twins, personality-driven matching, and agent-curated feeds
- **Economy** — native wallet infrastructure, token utility, and agent-mediated value exchange
OpenHermit provides the production runtime that makes this possible: durable state, sandboxed execution, multi-channel delivery, fleet management, and the access policy system that governs how agents act on behalf of users.
Amiko is agent-agnostic by design — bring your own agent or create one on the platform. Both paths lead to the same social graph, the same identity layer, and the same economic rails.
Learn more at [heyamiko.com](https://heyamiko.com).
---
## Contributing
OpenHermit is open source (MIT) and very much a work in progress. If the internal/external state split, the fleet operations model, or agents-as-services resonates with you — we'd love your help.
Issues, PRs, design discussions, channel adapters, skills, MCP integrations, docs, and war stories from running it in your own setup are all welcome.
- Open an issue: [github.com/HCF-S/openhermit/issues](https://github.com/HCF-S/openhermit/issues)
- Start a discussion: [github.com/HCF-S/openhermit/discussions](https://github.com/HCF-S/openhermit/discussions)
- Or just star the repo if you'd like to see where this goes
---
## License
MIT © [Amiko](https://heyamiko.com)