⛰️ Olympus

Your machine, your memory. Keep your private things private.

A private second brain on your own computer, for ChatGPT, Claude, Grok and any AI app that speaks MCP. On a Mac, and on Omarchy and Ubuntu 24.04 LTS Linux. Free and open source.

olympusplugin.ai · Get started · Olympus in ChatGPT · Claude, Grok and other apps · How it works · Sources · Security model · OpenClaw quickstart · Commands

🤖 Installing this as an AI agent for OpenClaw? Stop and read INSTALL_FOR_AGENTS.md before running any command — it is the step-by-step runbook you must follow, including exactly what to tell your operator after each step (your install report must end with the Step 1→2 invitation, never a bare status line). Managed git:, clawhub:, and npm-pack: installs intentionally hide their internal package path; use the installed olympus CLI and this packaged runbook rather than searching OpenClaw's managed storage. Setting up Olympus for ChatGPT, Claude or Grok instead? Follow Get started, not this runbook.

Frontier AI models are the smartest, but using them means sending your life to someone else's cloud. Local models keep it home, but they are a step behind, and the best ones need an expensive computer under your desk. Olympus gives you both, on the laptop you already have.

Every email, file and message is judged, one by one, as Personal, Private or Secret: on your own computer by default, or by Venice's private models if you choose a Venice-only setup.

• Personal: ChatGPT, Claude or Grok can read the parts that answer your question, and cite them.
• Private: only a model on your computer reads it (or Venice, if you turn it on); your AI app never gets the item itself. In ChatGPT, only you see the answer; other apps get an answer written from it, which can repeat what it says.
• Secret: passwords, recovery codes and keys. Once Olympus recognizes a Secret, it is never sent or answered from.

Labels are automatic and can be wrong.

Everything is stored and searched on your computer: the search index is built there, by a model small enough for a computer with 8 GB of memory. There is no hosted copy of your things, no account and no subscription. Olympus's publisher runs only a relay that passes your AI app's requests to your computer without storing them. Google sign-in happens on your computer, directly with Google. For hard private questions you can reach stronger models: Venice runs private models that, under Venice's own terms, keep nothing you send, and zkAPI, from OpenAnonymity and the Ethereum Foundation, hides who is asking (see Stronger models, privately).

you › "What did I commit to this week, across email and Telegram?"

agent › Three commitments: the tax documents to Maria by Friday
        (email, Tue), reviewing the lease amendment (email, Mon),
        and the Saturday ride logistics (Telegram, Wed). [3 citations]

Your agent saw a bounded, cited, policy-gated answer. It never saw your mailbox.

Get started

You need a Mac with Apple silicon (M1 or later) on macOS 13 Ventura or later, or an x86_64 computer running Omarchy or Ubuntu 24.04 LTS. Intel Macs, other Linux distributions and ARM Linux are not supported yet. You do not need OpenClaw, Tailscale, an API key or a new account.

1. Install Olympus on your computer. On a Mac, open Terminal and run:

curl -fsSL https://olympusplugin.ai/install.sh | sh

On Omarchy, first install the tools Olympus uses to read scans, photos and voice notes, then run the Linux installer:

sudo pacman -S --needed tesseract tesseract-data-eng poppler libvips libheif ffmpeg
curl -fsSL https://olympusplugin.ai/install-linux.sh | sh

On Ubuntu 24.04 LTS, install the prerequisites and run the Linux installer:

sudo apt install curl ca-certificates tar unzip coreutils findutils gawk grep diffutils procps util-linux systemd python3 tesseract-ocr tesseract-ocr-eng poppler-utils libvips-tools libheif1 ffmpeg
curl -fsSL https://olympusplugin.ai/install-linux.sh | sh

Each installer installs Olympus for your user only (the installer itself needs no administrator password; the package installation lines above do), checks every download against a pinned SHA-256 before using it, keeps Olympus running in the background (a LaunchAgent on a Mac, a systemd user service on Linux), adds the olympus command, and waits until the engine proves it is running. On Omarchy it also adds Olympus to the app launcher. Then Olympus downloads its built-in models once. Running it again updates or repairs the install; a failed update puts the previous version back. Details, folders and uninstall: https://olympusplugin.ai/install/.

2. Connect your AI. Each app has a short guide: ChatGPT · Claude · Grok and Grok Bot · Dots · Claude Code · Codex · OpenClaw · Hermes · any MCP app. In ChatGPT, Claude and Grok you add Olympus's address as a connector (in ChatGPT, in Developer mode until Olympus is listed in its directory), then approve it with one click on a page Olympus opens on your own computer: there is no pairing code and no account to create. Grok Bot then uses the Grok connection. Apps on your computer, such as Claude Code, Codex and Hermes, connect with one command.

3. Connect your sources and ask. Open the dashboard (in ChatGPT, ask "Show my Olympus dashboard"; on your computer, run olympus dashboard), connect Dropbox, Gmail or Google Drive, pick folders, and ask the way you normally would.

Olympus for ChatGPT

In ChatGPT, Olympus is a plugin. The Olympus engine runs on your computer and keeps your index there; ChatGPT asks it questions through one secure connection that you approve on that computer.

1. Add Olympus in ChatGPT

Install Olympus on your computer first (Get started). Olympus is on its way to ChatGPT's app directory. Until it is listed, add it yourself: in ChatGPT, open Plugins, choose +, then Add custom MCP server. Name it Olympus, choose OAuth, and paste https://mcp.olympusplugin.ai/mcp as the URL. If you don't see Add custom MCP server, turn on Developer mode in Settings first. OpenAI decides which accounts have Developer mode; in a work or school workspace, an admin may need to allow it. Step by step: https://olympusplugin.ai/install/chatgpt/. On Omarchy, ChatGPT may still say Olympus needs a Mac; it works on Omarchy too.

2. Approve ChatGPT on your computer

The first time ChatGPT connects to Olympus, an Olympus page opens in your computer's browser asking Connect ChatGPT to Olympus? Click Connect. (If a page saying Approve ChatGPT on your Mac comes first, click Continue on this Mac; the page says Mac on Omarchy too.) Clicking Connect on your own computer is how Olympus knows the computer is yours: there is no code to copy and no account to create. To check on it later, run olympus engine status in a terminal.

3. Connect Dropbox and choose folders

Open the Olympus dashboard in ChatGPT's sidebar (or ask "Show my Olympus dashboard"). Choose Connect next to Dropbox and sign in with your own Dropbox account; the sign-in link works once, for 10 minutes, from any device. Gmail and Google Drive connect the same way, except that you open their sign-in link on the computer where Olympus runs, not on a phone.

Then pick folders. For each folder choose:

• Full: Olympus reads the files and can answer from what they say.
• Names only: Olympus keeps the file names and details such as dates, not what the files say.
• Skip: Olympus leaves the folder alone.

A folder follows its parent's choice unless you change it; Mixed means some folders inside it are set differently. Choose Save and start, and indexing begins on your computer.

4. Tell Olympus what's private

Olympus sorts every item, one by one, into one of three tiers:

• Personal: everyday mail, notes and files. ChatGPT can see the parts that answer your question, and cites them.
• Private: things you keep to yourself, such as health, money or family matters. What they say is never sent to ChatGPT. Olympus answers from them on your computer, in the private answer panel (below).
• Secret: passwords, recovery codes and keys. What Olympus recognizes as a Secret never reaches a model and is never answered from. Recognition looks for common forms (private keys, API keys, labeled passwords) and can miss others.

Olympus decides on your computer. Once its built-in private model is ready, it checks every item it reads before treating it as Personal. Labels are automatic and can be wrong. In the dashboard, Privacy asks "What's private for you?": answer in your own words. Once you do, only the kinds of things you call private are Private: health or family matters you leave out, or say are fine to share, can reach ChatGPT. You can also mark folders, Gmail labels or senders as always Private.

5. Ask

Ask in ChatGPT the way you normally would: "What did Sam say about the lease?" ChatGPT searches Olympus and answers from your Personal items, with citations.

The private answer panel

When some of the items that match your question are Private, a small Olympus panel appears on its own under ChatGPT's reply: Private answer from your computer, marked Not sent to ChatGPT.

• Your computer writes the answer with its built-in private model (Qwen3.5, sized to its memory: 2B under 16 GB, 4B with more) and sends it to the panel sealed, so only the panel can open it. It goes around the ChatGPT conversation, not through it, and the Olympus relay cannot read it.
• On an Apple silicon Mac it usually arrives within a minute; on Omarchy it depends on the computer's processor and can take a few minutes. Ask for "all the details" and Olympus reads the whole documents instead; that can take a few minutes.
• Sources (closed at first) lists the items the answer used. Click one to open it on your computer, or in Dropbox on the web if it is not on this computer (ChatGPT opens that web link for you, so it sees the address).
• Hide folds the answer away and Show brings it back. An answer expires after 10 minutes; ask again for a new one.
• Ask follow-up questions in the chat as usual; each private answer appears in a new panel.

Your computer needs to be awake and online. The panel runs inside ChatGPT's page, so it relies on the panel code ChatGPT loads; the design states that limit plainly.

What ChatGPT can and can't see

ChatGPT can see:

• The parts of your Personal items that answer a question: excerpts, titles, dates and links.
• File names, including the names of some Private items. A file's name can be Personal while what it says is Private, so its name may show up in ChatGPT; its contents never do.
• That some items matching a question are Private: ChatGPT's model learns just that, with no titles and no content.
• Status: which sources are connected, how far indexing has got, and what needs you. The folder picker shows your folder names inside ChatGPT so you can choose; they pass through ChatGPT to reach the picker, but ChatGPT's model is not given them.

ChatGPT can't see:

• What your Private items say, or the private answers made from them.
• Secret items Olympus recognizes.
• Folders you skip, or what the files in Names-only folders say.
• API keys or passwords: you never enter one in ChatGPT. Optional extras such as a Venice account are set up on your computer.

The relay at mcp.olympusplugin.ai passes ChatGPT's requests to your computer. It stores and logs no questions, answers or tokens, but like any HTTPS service it handles them in transit. Full details: the ChatGPT design and the trust model.

Claude, Grok and other apps

Claude (web, desktop and phone) and Grok connect through the same Olympus relay as ChatGPT, with one click to approve on your computer, and Grok Bot uses the Grok connection; apps on your computer, such as Claude Code, Codex and Hermes, connect there directly. Guides: https://olympusplugin.ai/install/.

They search what is Personal, with citations, like ChatGPT. They never get your Private items themselves or the Secrets Olympus recognizes. For Private items they get at most an answer that the private model on your computer (or Venice, if you set it up) writes from them, checked for secrets, with each item's title, path, source and author. That answer can repeat what the items say. ChatGPT is stricter: it gets no answer from Private items at all, and shows one only in its private answer panel. When an agent is not given an answer, Olympus records the question and the model built into Olympus answers it on your computer; the agent is told only that the answer is in Olympus on your computer, with a link (olympus://open/private/ and a random identifier) that opens it there. The answer waits in the dashboard's Inbox (under On this computer), which Olympus gives no AI app a way to read, until you delete it, and Needs you counts the answers you have not read. More: Connect other agents.

Stronger models, privately

For hard private questions you can reach stronger models without giving yourself away. Both are off until you turn them on.

• Venice runs private models that, under Venice's own terms, keep nothing you send. Add your Venice key under Keys, on your computer. In Ask a question, a Venice model reads the passages of your files it needs, and is available only if your privacy settings allow Private items to go to Venice; a key alone does not allow it.
• zkAPI, from OpenAnonymity and the Ethereum Foundation, hides who is asking. It pays anonymously and sends the question over Tor (on unless you turn it off; without Tor, your network address shows): no account, no card, no name, just proof the question is paid for. In Strict mode your computer first rewrites the question into general questions; Standard, by default, takes out names, contact details and account numbers, or sends your words as written if you choose that. What you ask can still identify you. Turn on Anonymous answers under Keys on your computer, then ask from your agent or from Ask a question in the dashboard's Inbox; each question as sent, and its reply, stay in the Inbox on your computer until you delete them or the Inbox removes the oldest when full. Anonymous answers work on a Mac and on Omarchy and Ubuntu 24.04 LTS (x64), once Olympus's network isolation checks pass.

Uninstall

On a Mac, run curl -fsSL https://olympusplugin.ai/uninstall.sh | sh; on Linux, curl -fsSL https://olympusplugin.ai/uninstall-linux.sh | sh. Both stop Olympus and remove the app, the runtime and the olympus command, and keep your settings and data. To delete those too, first run olympus engine stop and olympus data delete --all. To only stop Olympus and remove its login item or service, run olympus engine uninstall.

The rest of this README covers how Olympus works and Olympus with OpenClaw, which stays supported and optional.

How it works

Olympus ingests your sources into local, trust-tiered indexes, then lets your frontier agent answer questions across all of them, with sensitive material handled only by lanes you approved: your own local models, privacy-approved private cloud inference, or not at all.

Every item you ingest is judged on its own as Personal, Private, or Secrets (plus Public on installs made before 1.0; internally S0-S5): its names are Personal unless something raises them, and its content is raised to Private or Secrets on evidence. Each tier has its own index and routes only to the model lanes your sovereignty policy allows:

Olympus uses Personal (ordinary personal and work content), Private (sensitive material), and Secrets (once recognized, never model input). There is no Public tier on any new install, ChatGPT or OpenClaw, whichever posture you choose: anything that would have been Public is Personal. Installs made before 1.0 keep a Public tier. See the tier mapping for the unchanged storage identifiers.

Four postures, chosen (and changeable) in setup — a config file, not a code fork:

“Private cloud only” describes Private-data handling: Venice answers Private questions. Search indexing for every tier uses EmbeddingGemma 2, Google's small embedding model, built into Olympus: it downloads once, runs on your computer, needs no account or key, and sends nothing anywhere. On a Mac it also reads pictures, so photos are searchable by what is in them. The embedding choices are built-in local EmbeddingGemma or Venice. Existing Gemini and local-server configurations remain compatible; see the embedding migration guide. Private content never goes to Gemini. Personal questions use OpenClaw's own configured default model.

One question searches every tier. When it finds Private evidence, the private lane your posture approved answers it and your agent receives only the checked, cited answer. Secrets Olympus recognizes are never searched by content: it can tell you where one lives, never what it says.

Some rules are not configurable, by design: secure content never routes to ordinary cloud models, recognized secrets are denied to every lane, and an exhausted policy chain refuses rather than silently downgrading.

Get started with OpenClaw

Using ChatGPT? See Olympus for ChatGPT instead; none of this section is needed.

You need OpenClaw 2026.7.1+ on a Node release OpenClaw itself supports (>=24.16.0 <25 or >=26.1.0 for OpenClaw 2026.9.5; check npm view openclaw engines — its installer refuses any other), plus Bun 1.2+. macOS or Linux.

Paste this into OpenClaw, or another agent with a terminal:

Install Olympus by reading https://raw.githubusercontent.com/jamiezigelbaum/Olympus/main/INSTALL_FOR_AGENTS.md and following it step by step. Walk me through setup and ask me about my data and privacy preferences.

Your agent downloads the designated pilot package, checks it automatically, and guides you through setup. You do not need to download files, find a checksum receipt, or have a GitHub account. The agent checks the SHA-256 and byte count against the designated candidate and GitHub's release metadata.

Olympus is also published on ClawHub (https://clawhub.ai/plugins/olympus, from 1.3.4), so openclaw plugins install clawhub:olympus installs the same release bytes. ClawHub can trail the newest GitHub release by a little. The agent install above and release qualification use the managed npm-pack: path — see docs/V0_4_RELEASE.md. A Git checkout or rebuilt package does not qualify the supplied candidate.

--accept-capabilities is required by OpenClaw 2026.8.1+: an agent-driven install is non-TTY, and without the flag the host exits 1 asking for capability consent it cannot prompt for. Omit the flag on 2026.7.1, which does not define it.

The agent checks prerequisites, installs the plugin, talks through what is private for you (you save it later in the dashboard's Privacy section), asks for your privacy posture and your approval of the private classifier, verifies the worker, restarts the gateway, and hands you the dashboard. You paste your model keys into its Models section yourself; the agent never collects them. That completes base installation. Then you can choose a source in the dashboard or leave Olympus ready for later; no source is selected for you. Once your chosen source is ready, the agent checks a first answer and its citations.

Prefer to drive it yourself? Follow docs/QUICKSTART.md, starting with the automatic download and existing-install checks. Its install command for a clean machine is:

openclaw plugins install npm-pack:/absolute/path/to/olympus-1.4.1.tgz --force --accept-capabilities
# On OpenClaw 2026.7.1, omit both flags for a clean install.
# On newer hosts, --force also overwrites an existing plugin: check first.
openclaw plugins enable olympus
OLYMPUS_ROOT="$(openclaw plugins inspect olympus --json | jq -r .plugin.rootDir)"
OLYMPUS_BIN="$OLYMPUS_ROOT/bin/olympus"
"$OLYMPUS_BIN" doctor

Continue the quickstart to describe your data, choose a privacy posture, connect model credentials, and verify base activation before choosing your sources in the dashboard and checking the first cited answer. Setup installs an owner-only ~/.local/bin/olympus launcher on Linux and macOS and keeps it aligned with worker upgrades and rollback. Existing user or standalone launchers are preserved. If ~/.local/bin is missing from your login PATH, run "$HOME/.local/bin/olympus" dashboard on the machine running Olympus (on a server, use that server). Setup reports the full command, including a fallback when an existing launcher belongs to another installer.

This candidate artifact includes native Control UI support. On OpenClaw 2026.9.2, use Olympus in the Control UI sidebar to connect sources, choose scope, and follow ingestion. Enable Settings → Labs → Custom plugin UI (the Gateway restarts itself to apply it on OpenClaw 2026.9.5; run no second restart), then reload the browser. OpenClaw currently makes this integration experimental and requires its own Gateway's Control UI over HTTPS or localhost. Olympus uses your signed-in OpenClaw permissions; its worker token stays on the server. The standalone olympus dashboard command remains available for older hosts, when custom plugin UI is off, and for direct access when needed. See the dashboard guide.

For Venice embeddings, Venice accounts/API credit, local answer models, or legacy embedding migration, use the agent-led model setup guide. It explains the separate Private and Personal routes, registered embedding defaults, and custom-model requirements before you connect keys or restart. In the dashboard you paste each key into its Models card. For a headless setup, choose how to supply each key: paste it yourself from your password manager's website into the silent terminal input, or authorize an exact named-item read through an already authenticated manager CLI. The manual route needs no password-manager desktop app or CLI; a 1Password op:// reference needs authenticated op access for an agent fetch.

On OpenClaw 2026.9.2, use Olympus in the Control UI sidebar when the installed Olympus artifact includes native Control UI support and Custom plugin UI is enabled. Host version alone is not proof of support. Otherwise the quickstart uses the retained standalone dashboard; it explains the different authentication requirements for each route.

Supported sources

Model-lane credentials such as Venice are security-preset prerequisites, not source connectors. Domain-agent imports and other source families are outside the 1.0 public roster.

Tokens live in an encrypted local secret store by default, or in an explicitly configured OS secret store when supported, never in plain text. Ingestion, classification, and indexing all run in one supervised local worker — olympus worker install makes it start on login. All seven declared sources sync through the canonical connector-store runtime; Olympus supports one connected account per provider. After base activation, dashboard onboarding guides the source you choose from credentials and scope through initial sync and cited-answer readiness, with the next supported action on blocked or degraded sources. The seven-source roster is not an installation checklist. Clean-install and real-provider qualification remain Slice 4 work, not alternate ingestion paths.

On an existing live OpenClaw install, treat plugin install/enable, gateway restart, worker service changes, and plugin refresh as live-system changes: validate with openclaw config validate and openclaw doctor --lint, then restart only through openclaw gateway restart or the documented ops refresh script. The repo docs and commands should never require raw edits to OpenClaw runtime config.

The security model

Olympus is built on the assumption that the boundary is the product:

• Local-first ingestion. Raw source content lands in local SQLite indexes before any model sees anything. Classification happens on your computer unless you choose Venice for it.
• Sanitized handoffs. Agents receive answer text, excerpts of Personal items, citations, and audit counts — typed wire shapes with raw_source_exposed: false, re-validated client-side. Embeddings, cursors, and tokens never cross. Private items reach an agent only as answers written from them, which can repeat what they say; ChatGPT's model gets none.
• A hard membrane. Any evidence pack containing sensitive material is barred from standard cloud lanes by two independent checks — policy validation at config load, and a runtime re-check on every request.
• Fail closed, everywhere. No approved lane for a tier means a refusal with an honest error — never a silent fallback to a less-trusted model.
• Authenticated by default. The worker binds loopback with bearer-token auth on every route; new installs generate the token during setup. Browser controls exchange it for a short-lived HttpOnly local session protected by same-origin and CSRF checks; the bearer is never stored by the page.
• A real exit. olympus data export gives you your data; olympus data delete --all verifiably removes indexes, embeddings, configs, tokens, and service units.

The onboarding flow, end to end:

Command reference

olympus worker install and upgrade atomically reconcile the generated service unit and owner-only worker environment before activation. Interrupted install or upgrade transactions recover the previous managed files when the same command is rerun. status reports the observed service and transaction state; it never guesses that an unreadable manager or non-regular managed path is healthy. uninstall removes only the supervised service and intentionally retains credentials, source configuration, and indexed data. Use the separate data lifecycle when those bytes should also be removed.

Credential, scope, pairing, Disconnect, partial-sync, and source-dependency repairs do not normally require a worker restart. Follow the source card or olympus doctor recovery action instead.

Dashboard Disconnect stops new scheduled and manual reads, removes the selected local account grant, and keeps indexed data and reusable developer-app registration. It does not revoke provider-side access; the card links the provider revocation page. Local data deletion remains the deliberate CLI-only flow: preview olympus data delete --source <id> --dry-run, Disconnect the source, then run the command without --dry-run.

Telegram and WhatsApp are paired sessions rather than broker grants, so their cards carry Unpair instead: it stops new reads and deletes this computer's pairing session, keeping indexed messages and captured media, and leaves the linked device in place at the provider for you to remove there.

Connect other agents

Open the dashboard on this computer (olympus dashboard). Under On this computer, choose Open next to Agents, then press Connect an agent. Pick the agent you use and follow its steps; every address and snippet has a copy button.

• ChatGPT: add Olympus in ChatGPT (Developer mode until it is listed; see Olympus for ChatGPT), then click Connect on the page that opens in this computer's browser.
• Claude Code and Codex on this computer: copy the prompt (or the olympus serve command) into the agent. This works without remote access.
• Claude (web, desktop and phone) and Grok (grok.com): add Olympus as a connector with the address the panel shows, then click Connect on the page that opens in this computer's browser.
• Grok Bot: nothing to add. Once Olympus is connected on grok.com, Grok Bot uses that connection, with the same xAI account.
• Agent on another account of this computer: for OpenClaw, Hermes or an MCP client running under another account. Name the agent and choose Create key, then copy its setup line into the agent. It connects to this computer only, without reading your Olympus files or signing in.
• Agent on my tailnet: for an agent on your Tailscale network. Keep Tailscale running on both computers, then create a key and copy the setup line. If Tailscale is not running here, the page says so.
• The Grok API and other agents in the cloud need a tunnel you run yourself (below), unless the agent can use one of the private choices above.

The private choices use a different key for each named agent. Each address starts listening only while it has an active key; revoking its last key stops it. The connection list shows which route each key uses. Keys and setup lines are shown once, so copy them before choosing Done. These connections use the same remote tools and privacy rules as the relay, with no question limits. They work independently of the remote access switch.

Agents in the cloud reach Olympus only when remote access is on; the Agents page says whether it is. To turn it on, press Turn on remote access there. Olympus then connects through the Olympus relay (mcp.olympusplugin.ai): you create no account and run no tunnel. The relay passes requests between your agents and this computer; it stores and logs no questions, answers or tokens, but like any HTTPS endpoint it handles them in transit. The relay connects ChatGPT, Claude, Grok (grok.com) and Grok Bot, and approving one happens on this computer only, with one click on a page Olympus serves locally; there is no pairing code. A Claude connector approved on this computer then works in Claude on your phone too. Turn off remote access turns it off again. If the relay cannot be reached, the page says Olympus relay unavailable and Olympus keeps retrying on its own; agents on this computer are unaffected.

Advanced: to use a tunnel you run yourself instead of the relay, set plugins.entries.olympus.config.remote.publicBaseUrl to its https origin (openclaw config set …) and turn remote access on. Through your own tunnel, Claude, ChatGPT and Grok approve Olympus on its approval page with a one-time pairing code from the Agents page, and the Grok API takes an address plus a key from Create key (shown once).

Every connected agent gets the same privacy rules as your OpenClaw agent. It never gets Private items themselves or the Secrets Olympus recognizes. For Private items it receives only answers that Venice or a local model wrote from them, which can repeat what they say, with each item's title, path, source and author. The Agents page lists connected agents with when each was added and last used, and Revoke removes one. olympus connections list|revoke|pair|add does the same from a terminal.

Each agent's steps end with a short instruction that tells it when to ask Olympus. Paste it into Claude project instructions, Grok's custom instructions or Grok Bot's skills, or load the packaged Agent Skills folder integrations/agent-skills/ask-olympus/.

Using Olympus from other agents (MCP)

olympus serve exposes the same sanitized read operations — source_answer, source_answer_result, source_index_status, capability-gated source_index_search, ask_anonymously, and archive_import — to any MCP-capable agent. The supported Hermes configuration narrows that server to exactly source_answer, source_answer_result, source_index_status and ask_anonymously through its per-server tool allowlist.

archive_import (start, status, remove) lets an agent ask to import an old export the user wants searched, such as a mail .mbox from Google Takeout or Apple Mail. It only asks: start stores the path as text without opening it and returns needs_approval; the user approves on the Olympus dashboard on their computer (Sources, Archives), and remove works the same way. status reports each archive's progress, counts, date range and gaps, never message text. It is on the native, MCP and CLI surfaces only, not on Hermes, remote MCP or the ChatGPT relay.

ask_anonymously sends one typed question to a frontier model through zkAPI (paid per question, anonymously, from the user's own balance; what the question says can still identify them). The first call with no level returns needs_choice: the agent asks the user once for Strict (their own model rewrites the question into general questions before it leaves) or Standard (their words, prepared as written, lightly cleaned up or by their saved instruction), then calls again with level and remember: true. Only the question goes out, never documents, and never to the provider that hosts the agent asking: a question from ChatGPT goes to Claude Sonnet by default, one from Claude Code or Claude Desktop to an OpenAI model, and a one-off model from the agent's own provider is refused. Setup (the zkAPI route, the writer model, the zkAPI models for questions from ChatGPT and from Claude, the spend limit) lives on the dashboard's Anonymous answers card, under Keys. Questions are asked through an agent, olympus ask, or Ask a question in the dashboard's Inbox; each question as sent, and its reply, are kept in the Inbox on your computer until you delete them or the Inbox removes the oldest when full. See docs/design/private-answers.md.

MCP clients cap how long one tool call may run, and an answer can take several minutes. So over MCP (local olympus serve and remote /mcp alike) and the OpenAPI tool paths, a source_answer that is not done by a hand-off threshold returns {"status": "working", "job_id": …} and keeps running. The threshold is 200 seconds for remote agents (OLYMPUS_SOURCE_ANSWER_HANDOFF_MS) and 45 seconds for local olympus serve (OLYMPUS_SOURCE_ANSWER_STDIO_HANDOFF_MS, under Codex's 60-second default tool timeout); both are capped at 230000. source_answer_result with that job_id returns the answer, the same error the call would have raised, or working again after waiting up to a minute (never longer than the threshold). A slow ask_anonymously (a Tor route can take minutes) hands off the same way and is collected with the same tool. A job that runs past 20 minutes is stopped with source_answer_deadline, and revoking a connection drops its jobs.

Remote answers share the analyst with your own assistant's answers; there is no priority lane. So the worker runs at most two remote answers at once (the analyst's lane count; OLYMPUS_SOURCE_ANSWER_MAX_RUNNING raises it for a multi-lane analyst) and refuses more with source_answer_busy, which bounds how much remote work can ever queue ahead of your own question. A caller that disconnects, or a job that reaches its deadline, stops the analyst call itself and frees the lane. A job id is readable only by the connection (or olympus serve process) that asked, lives in memory, and expires 15 minutes after its answer is ready.

For Hermes Agent, resolve the installed plugin first:

openclaw plugins inspect olympus --json

Copy plugin.rootDir from that content-free response, append /bin/olympus, and pass the resulting absolute executable path to Hermes:

hermes mcp add olympus --command /absolute/managed/olympus/bin/olympus --args serve
hermes mcp test olympus

Make the olympus server entry in ~/.hermes/config.yaml match the packaged config/hermes/olympus.mcp.yaml:

mcp_servers:
  olympus:
    command: <absolute-managed-plugin-root>/bin/olympus
    args: [serve]
    tools:
      include: [source_answer, source_answer_result, source_index_status, ask_anonymously]
      prompts: false
      resources: false

Restart Hermes or run /reload-mcp, then verify that the discovered registered names are exactly mcp_olympus_source_answer, mcp_olympus_source_answer_result, mcp_olympus_source_index_status and mcp_olympus_ask_anonymously. Invoke the discovered names rather than hard-coding them. A cited-answer round trip uses the discovered mcp_olympus_source_answer; source_watch_* remains OpenClaw-only.

The optional packaged Hermes skill is integrations/hermes/ask-sources/SKILL.md. It is also published on ClawHub, which Hermes's skills hub searches, so hermes skills install clawhub/olympus-hermes installs it. Or copy its directory to ~/.hermes/skills/ask-sources/ or expose the parent with skills.external_dirs; it declares the same four tools and no fallback access.

No hermes://mcp/install link is published: current Hermes upstream documents custom MCP installation through hermes mcp add but does not document that URI handler.

There are no operator agent tools. Index maintenance — sync, extraction, embedding, retries — is the worker's own scheduler, and what a person drives by hand goes through the olympus CLI. The whole agent tool surface is thirteen tools — argus_ping, argus_list_models, argus_complete, source_answer, source_answer_result (MCP and remote only), source_index_status, source_index_search, source_watch_create, source_watches, source_watch_cancel, olympus_doctor, ask_anonymously, archive_import — declared in src/core/public-surface.ts; Hermes over MCP sees the subset above, and worker bearer auth is enforced throughout.

Operational note for connector-store corpora: a source with an account folder (Gmail, Drive, Dropbox, local files, Readwise, X, Telegram, WhatsApp, imported archives) keeps its stores where Olympus places them, and its OLYMPUS_SOURCE_INDEX_CONNECTOR_STORES_JSON entry names no path: it only adds a principal to the store its lane opened. A store only the mount list places (Reflect, Roam) is mounted with the same corpus id, family, trustDomain and dbPath its sync script writes, and that corpus id must be in the corpus registry — a store mounted for an unregistered corpus fills but is never read.

Chat-family mounts must also declare their exact principal identity with principalProvider and principalAccountScope (for example, whatsapp and personal). Structured scopes use the stored account scope as their prefix, so the shipped WhatsApp identity is personal:chat:<id> while Telegram remains telegram.personal:chat:<id>. A chat mount without both fields remains available for non-chat_scope reads, but chat_scope filtering fails closed as not configured; mounted rows never establish account authority.

The migration-era indexes, replay/import utilities, read-authority switches, and retired store writers were removed in Slice 2. Canonical connector stores are now the only source read/write authority. Compatibility aliases that remain at input boundaries do not select another runtime or store.

Development

Olympus is TypeScript on Bun: one plugin package, stable versioned architecture contracts (SourceConnector / EvidencePack / Analyst — only the connector is per-source), and 3,400+ tests including an architecture guard that keeps per-source answer logic out of the shared spine.

This release tarball contains the built plugin runtime. Development verification runs from the source repository, where the full src/, scripts/, test/, and eval/ tree is present.

Use focused local proof while editing; GitHub requires each substantive CI lane and runs the complete suite in parallel before merge. See the canonical engineering process.

Deep-dive docs: contracts and architecture · trust model · sovereignty config · uninstall

License

MIT — free for any use, including commercial. Olympus is provided as is, without warranty; keep your own backups. Olympus is published by Open Coordination Unlimited, Inc. Its hosted relay is governed by the Terms of Service and Privacy Policy. Third-party data notices: docs/THIRD_PARTY_DATA.md.

Want to suggest a change? See CONTRIBUTING.md: fork, branch, open a PR. Read access doesn't include direct pushes — everything lands through review.