Skip to content

Set up Signet

signet setup creates your workspace, starts the daemon, and opens onboarding: nine screens that name your agent, connect the tools you use, and choose how memory runs.

Run it again whenever you want to change those choices. It picks up your current settings and resets nothing.

Terminal window
signet setup
Workspace: /Users/you/.agents
Continue setup: http://127.0.0.1:3850/#setup
Connect your provider, select a memory model, and test the connection.

Setup needs an interactive terminal. On a new workspace it creates:

  • agent.yaml, the workspace configuration
  • data/signet.db, the memory database
  • folders for files, transcripts, skills, and secrets, plus a .gitignore

Then it starts the daemon on 127.0.0.1:3850 and opens onboarding in the desktop app if it is installed, otherwise in your browser. If nothing opens, paste the printed URL. signet configure opens the same flow.

A new workspace starts with these defaults. Onboarding changes them.

Setting Default
Identity files Off. Your harnesses keep their own instructions.
Search Keyword only. No embedding model.
Dreaming On. It waits until a model provider is connected.
Telemetry Off.

Onboarding is a full-page view in the dashboard. Three screens shape how Signet works: Identity, Harnesses, and Connection. The rest are optional or informational, except First memory, which proves the loop works.

Screen You
1 Welcome Read and continue.
2 How it works Read and continue.
3 Identity Name your agent. Required.
4 Harnesses Pick the agents you use.
5 Connection Connect a model, or set it up later.
6 Sources Import notes or files. Optional.
7 Search Choose keyword or semantic search.
8 First memory Save one memory and recall it. Required.
9 Summary Review and open Signet.
The Identity screen with Agent name, Your name, and Identity preset fields.The Identity screen with Agent name, Your name, and Identity preset fields.
  • Agent name required

    Your agent’s display name, up to 80 characters. It appears in the dashboard and in identity file templates.

  • Your name optional

    Fills the name lines in USER.md and AGENTS.md when Signet manages identity files.

  • Identity preset default: Keep identity in my existing tools

    Whether Signet writes and injects identity files, and which ones.

Keep the default if your harness already has the instructions you want. Choose a preset to let Signet manage identity files across every connected agent. Each file opens in an editor before you save.

Preset Files Signet writes
Keep identity in my existing tools None
Minimal · instructions only AGENTS.md
Hermes · personality + instructions SOUL.md, AGENTS.md
OpenClaw · full identity AGENTS.md, SOUL.md, IDENTITY.md, USER.md, HEARTBEAT.md, BOOTSTRAP.md
Custom · choose your files The files you check

Identity files never change the memory database or recall.

The Harnesses screen with Claude Code selected among eleven agents.The Harnesses screen with Claude Code selected among eleven agents.

Select every agent you use: Claude Code, Codex, Hermes Agent, OpenCode, OpenClaw, Gemini, pi, Oh My Pi, Kimi, Muse Code, or ForgeCode. Connect agents installs each agent’s Signet integration on the machine running the daemon. You can skip this screen and run signet connect later.

  • Agents only pick up the hooks in sessions started after connecting.
  • Deselecting an agent here does not disconnect it.
  • OpenClaw needs its plugin package first. If it is missing, run signet setup --non-interactive --harness openclaw, then retry.
  • Hermes profiles other than the default connect from the terminal: signet connect hermes --profile <name>.
The Connection screen with ChatGPT, Anthropic, and Local model options, a provider list, and the Dreaming switch.The Connection screen with ChatGPT, Anthropic, and Local model options, a provider list, and the Dreaming switch.

Signet uses a model to extract memories from your sessions and to run Dreaming. Choose one:

  • ChatGPT / Codex or Anthropic (Claude): sign in with your subscription.
  • Local model: an OpenAI-compatible server on this machine, such as Ollama or LM Studio. The address must be a loopback address; the default is http://127.0.0.1:1234/v1.
  • Use another provider: any provider in the catalog. Providers that support sign-in use it; the rest take an API key, which Signet stores as a secret.

Then pick a memory model and select Test and enable memory. Signet saves the configuration, sends the model a short test prompt, and starts automatic memory processing. If memory processing is frozen or in shadow mode in Settings, enabling stops and tells you to review those controls.

Dreaming is on by default and saves as soon as you toggle it.

Set up later skips the connection. Saving and recalling memories still work, but signet status and Home show Memory is paused until you connect a provider. Connect one later under Settings → Inference. Dreaming stops waiting within about 30 seconds and runs once there is enough to process.

The Sources screen with Obsidian, Agent transcripts, and Files and documents options.The Sources screen with Obsidian, Agent transcripts, and Files and documents options.

Bring in context you already have. Each option is optional and you can add more than one.

  • Obsidian connects a vault by its absolute path and indexes it.
  • Agent transcripts imports conversation exports into an agent. See transcript imports.
  • Files and documents imports reference material.

Imports keep running after you continue. Follow their progress, errors, and retries under Sources.

The Search screen with Keyword search only selected.The Search screen with Keyword search only selected.

Keyword search finds memories that share words with your query. An embedding model also finds memories by meaning.

Option Model Runs
Built-in nomic-embed-text-v1.5 On this machine. Recommended. The first use is slower while the model loads.
Ollama nomic-embed-text Your Ollama server, http://localhost:11434
llama.cpp nomic-embed-text-v1.5 Your llama.cpp server, http://localhost:8080
OpenAI text-embedding-3-small OpenAI, with an API key stored as a secret
Keyword search only None Default

Keep current settings continues without changes. Change this later under Settings → Inference → Embeddings; switching models rebuilds the search index and leaves your memories intact.

The First memory screen after saving and recalling a memory about answer style.The First memory screen after saving and recalling a memory about answer style.

Write one thing your agents should know, up to 240 characters. Remember this saves it as a private memory for your agent, and Recall it searches for it. You continue once recall finds it. Retrying never saves it twice.

The Summary screen listing the agent, connected tools, background memory, search, sources, and first memory.The Summary screen listing the agent, connected tools, background memory, search, sources, and first memory.

The summary shows what onboarding configured and which checks ran. Open Signet goes to Home. Start a new session in a connected agent and ask what it knows about you.

signet setup on an existing workspace starts the daemon and reopens onboarding with your current settings filled in. It does not reset anything.

  • Onboarding remembers your place in this browser, up to the Connection screen. The connection test always runs again.
  • The saved place includes your draft first memory but never credentials.
  • Leaving before the summary asks for confirmation. Settings you already saved stay saved.
  • Passing configuration flags to an existing workspace requires --non-interactive.

For servers and scripts, run setup without prompts. These flags apply to a new workspace.

Terminal window
signet setup --non-interactive \
--name "My Agent" \
--harness claude-code \
--embedding-provider native \
--extraction-provider none

--extraction-provider none leaves the model connection for later under Settings → Inference. Headless setup also starts the daemon and, with built-in embeddings, loads the model.

  • --non-interactive

    Apply flags without prompts. Also works without a terminal.

  • -p, --path path

    Workspace directory. Defaults to ~/.agents.

  • --name, --description text · default My Agent, Personal AI assistant

    The agent’s display name and description.

  • --harness id · repeatable

    claude-code, codex, hermes-agent, opencode, openclaw, gemini, pi, oh-my-pi, kimi, muse-code, or forge.

  • --identity-mode managed, off · default managed

    Headless setup manages identity files by default. Pass off to match interactive setup.

  • --identity-preset minimal, hermes, openclaw, custom · default minimal

    Which identity files to write when managed.

  • --embedding-provider native, ollama, openai, none · default native

    Search model. --embedding-model overrides the model where the provider allows it.

  • --extraction-provider provider · default detected

    Model for memory extraction: claude-code, codex, opencode, ollama, llama-cpp, openrouter, openai-compatible, acpx, or none. --extraction-model and --extraction-endpoint refine it.

  • --obsidian-source path[::name] · repeatable

    Connect an Obsidian vault.

  • --agent name:policy[:group] · repeatable

    Create additional agents. Policy is isolated, shared, or group.

  • --remote-url origin

    Use a daemon on another machine. See Remote daemon.

  • --network-mode localhost, tailscale · default localhost

    Where the daemon listens.

  • --enable-dreaming

    Turn Dreaming on. New workspaces have it on already.

  • --skip-git

    Skip initializing Git and committing the new workspace.

  • --open-dashboard

    Open the dashboard when setup finishes.

The CLI reference lists every flag.

A plan is the full set of setup choices as JSON. It suits provisioning tools that already know every answer.

Terminal window
signet setup --non-interactive --name "My Agent" --harness claude-code --dry-run > plan.json
signet setup --file plan.json
  • --dry-run prints the resolved plan for a new workspace without applying it.
  • --file and --json apply a plan, and only to a new workspace.
  • --schema prints the plan’s JSON Schema.

--remote-url https://signet.example.com points the workspace at a daemon on another machine instead of starting one locally. The value is a bare origin, with no path or credentials.

  • Interactive setup requires the remote daemon to respond before it opens onboarding.
  • Headless setup records the URL without checking it.
  • Harnesses connected from onboarding and Obsidian paths refer to the daemon’s machine. To connect an agent on your own machine, run signet connect there.
Symptom Fix
signet setup is interactive and requires a TTY Run it in a terminal, or use headless setup.
Use --non-interactive to apply configuration flags to an existing workspace Add --non-interactive, or run signet setup alone and change settings in onboarding.
Another workspace is running at this address Stop the other daemon, or select its workspace.
Signet is not ready Run signet doctor, then retry setup.
Onboarding did not open Open http://127.0.0.1:3850/#setup.
Could not load the daemon or provider catalog Check signet status, then select Retry.
OpenClaw needs its Signet plugin package Run signet setup --non-interactive --harness openclaw, then retry.
Memory is still paused, frozen, or in shadow mode Review memory controls in Settings, then test again.