Runtimes

A runtime adapter wires one agent CLI into the shared brain: it injects relevant knowledge into prompts going in, and captures learnings coming out. digital-me setup installs every adapter it detects; this page is the per-runtime reference for what gets installed and how to verify it.

All adapters are installed idempotently — re-running merges into your existing settings and never clobbers other hooks. Every adapter talks to the same endpoint: brain-host, the hub below, which setup installs before any adapter so each registration carries its URL and token-file path.

brain-host (the hub)

digital-me install --runtime brain-host       # setup does this first
digital-me service brain-host status

The brain itself: one always-on Node service on 127.0.0.1:18791 serving memory_search / memory_get / wiki from a retriever over your wiki/ + tastes/ (hybrid full-text + embeddings, incremental re-index every 30 min) and the brain-orchestrator tools (tasks, agent_identify, learning_capture, traces_record, traces_query, m1_event_record, m1_score) on one bearer-gated POST /tools/invoke wire. It owns brain.db and the scheduler tick that fires your workflows. Installed as a launchd (macOS) / systemd --user (Linux) service so it survives reboots.

What lands:

• ~/.local/share/digital-me/brain-host — stable install link (the service's working directory)
• ~/digital-me/.data/brain-host.token — bearer token (mode 600); adapters carry this path, never the secret
• ~/digital-me/.data/retrieval.db — the index; ~/digital-me/.data/brain.db — goals, tasks, traces, learnings
• ~/digital-me/.data/.env — where you put GEMINI_API_KEY; the service loads it and every worker it dispatches inherits it

Verify: curl -s http://127.0.0.1:18791/health reports the index size, the scheduler state and which brain.db it opened.

Claude Code

digital-me install --runtime claude-code

What lands:

• ~/.claude/hooks/* — dm_*.sh lifecycle hooks (UserPromptSubmit injects matching wiki knowledge; Stop captures session learnings)
• ~/.claude/skills/digital-me/ — the protocol skill
• merged settings.json — hook registrations alongside whatever you already had, plus the openclaw-brain MCP server entry (the name is historical — it is brain-mcp-proxy forwarding to brain-host)

Verify: start a session and ask about a topic you know is in your wiki — the prompt context will show a [Digital Me] injection block.

Codex

digital-me install --runtime codex

What lands:

• ~/.codex/CODEX.md — protocol instructions
• the openclaw-brain MCP entry in ~/.codex/config.toml (historical name; brain-mcp-proxy stdio↔HTTP transport to brain-host)
• ~/.codex/hooks/* wired through ~/.codex/hooks.json — UserPromptSubmit / Stop / PreToolUse, with M1 application-rate tracking

Hermes

digital-me install --runtime hermes

What lands: ~/.hermes/SOUL.md gains the digital-me protocol section, the recall plugin injects matching knowledge at prompt time, and the MCP stanza in ~/.hermes/config.yaml registers the brain tools.

openclaw (optional)

digital-me install --runtime openclaw

Only if you also run openclaw agents — it is a runtime like the others, not a prerequisite. The install materializes an additive plugin overlay (digital-me-brain + digital-me-recall) into your openclaw extensions directory — stock openclaw plus an overlay, no fork, no rebase. The overlay registers the brain tools (tasks, agent_identify, learning_capture, traces_record, traces_query, m1_event_record, m1_score) with the gateway, injects wiki knowledge at prompt time, and hosts the proactive-learning rule engine. It opens the same brain.db as brain-host and leaves the scheduler tick to brain-host whenever brain-host is installed (exactly one process ticks a brain).

Keep openclaw current without losing the overlay:

digital-me update --runtime openclaw          # update to latest mature tag
digital-me deploy                             # merged-in-git → verified-live

Dashboard (optional service)

digital-me install --runtime dashboard        # + always-on service
digital-me service dashboard status

A local viewer over your accumulated state — sessions, entries, application rate, workflows, feed. Installed as a launchd (macOS) / systemd --user (Linux) service so it survives reboots; --no-service skips that.

Dream-cycle (optional, Python)

The nightly distillation pipeline that turns raw captured learnings into clean wiki entries. Requires Python ≥ 3.11 in a venv (Homebrew/Debian enforce PEP 668 — digital-me doctor prints the exact recipe for your Python).

It ships on PyPI as digital-me-dream-cycle:

python3 -m venv ~/.venvs/dream-cycle
~/.venvs/dream-cycle/bin/pip install digital-me-dream-cycle
digital-me dream-cycle                        # run it manually

Working on the pipeline itself? Install it editable from the repo instead: pip install -e "packages/services/dream-cycle[dev]".

The nightly schedule (dream-cycle-nightly, 3am) is registered with brain-host's orchestrator at install time — no manual cron needed. The LLM key comes from ~/digital-me/.data/.env (engine: standalone, the default config.yaml); digital-me doctor checks it there.

Digest (optional, Python)

digital-me install --runtime digest

The 07:00 activity digest — what every agent did yesterday, what the dream-cycle distilled — posted to Discord. Delivery is a Discord webhook (no gateway needed): write the webhook URL to ~/digital-me/.data/digest-webhook.url (mode 600). Machines that still run an openclaw gateway can keep digest.delivery: openclaw in config.yaml to send through openclaw message send instead.

Dispatching work to your CLIs

The brain can also drive your CLIs: workflows dispatch steps to claude / codex through cli_exec_aliases in your config.yaml. setup pre-populates aliases for the CLIs it detects; see the Configuration reference for the shape.