Sign up and create one vault
Create your account, choose a personal or engineering starter, and confirm the active vault on Home.
Customer guide · about seven focused minutes
From one source to reusable evidence.
Create a vault, connect a supported agent, compile one small source, inspect the citation, and make the first trust decision yourself.
Use a disposable example source with no credentials or customer data. The snippets below contain placeholders only.
- Raw sourceUpload once
- Cited pageInspect the claim
- Human reviewPromote deliberately
- Agent reuseAsk again later
Know where each decision lives.
Helios is a set of inspection surfaces around one vault. Click a surface to learn its job; arrow keys move between tabs.
Home
Your first-value checklist.
Home tracks the real activation loop: source stored, citation visible, agent reused, and human trust decided. Its Ask form returns the best evidence section; it is not a chat transcript.
Reader
Read the page and challenge its source.
Open a page from Index, Search, Graph, or Pulse. Reader shows rendered Markdown, provenance links, trust state, table of contents, and backlinks. Follow an internal raw citation before promotion.
Pulse
See what needs attention.
Pulse separates uncompiled raw sources from compiled seedlings awaiting human review. It also reports validation, search freshness, backup cadence, and available capabilities.
Ask + Search
Find evidence; do not invent an answer.
Ask deterministically selects the strongest matching section and exposes its file and scores. Search may show a semantic badge when the capability is ready or a lexical/fallback badge when semantic infrastructure is unavailable.
Graph + Stack
Move from a claim to its neighborhood.
Graph visualizes wiki links; System Graph adds raw-source and operating relationships. Stack groups ideas, tools, and knowledge, with tours for source, question, and trust paths.
Log
Follow the operating record.
Log is the append-only journal of vault actions. Use it to confirm hosted compilation results, page writes, trust requests, and export activity without guessing from UI timing.
Settings + Export
Control access and prove the exit path.
Mint or revoke scoped keys, inspect the generated agent contract, import Markdown, configure privacy, and export a tar.gz containing reconstructed Markdown pages, byte-faithful visible raw files, logs, manifests, and checksums.
Complete one honest happy path.
Do not begin with a bulk import. A small source makes every boundary and failure state easy to see.
Add a small text source
Use Home to upload a source such as this harmless runbook. Upload stores raw evidence; it does not guarantee a compiled page.
oauth-runbook.md · example only
# OAuth rotation
Rotate signing keys every 90 days.
Record the new key id in the change log.
Revoke the previous key after the overlap window.Mint one scoped agent key
In Settings → API keys, choose principal agent, role agent, a recognizable client name, and an expiry. Copy the hk_… value once into your local environment.
Connect a supported client
Use Codex, Claude Code, or GitHub Copilot CLI. The quick configurations are below; the MCP guide is canonical.
Verify scope before writing
Call vault_list, read vault://VAULT_ID/AGENTS.md, then call vault_new_sources. The contract tells the agent how this vault handles ownership, privacy, provenance, and trust.
Compile by the plan you actually have
Hobby: call vault_source_read, separate extracted from inferred claims, and write cited seedlings with vault_remember. Pro or Team: call vault_ingest_source; a 202 started response means asynchronous work was accepted, not completed.
Find the claim and inspect its citation
Call vault_query with a concrete question, open the returned page in Reader, then follow its raw citation. Citations must use the returned raw id, not merely the filename.
Make the human trust decision
The agent may call vault_propose_promotion, but it cannot promote. In Reader or Pulse, compare the seedling with raw evidence and promote only if the page holds up.
Reuse the knowledge
Start a later or second supported agent session, load the same vault contract, and ask the original question. Success is a cited retrieval—not the agent remembering a chat.
Never paste a real key into a source, prompt, screenshot, or guide.
The examples use placeholders. If a key may have leaked, revoke it in Settings, mint a replacement with an expiry, update the client environment, and restart the client.
One endpoint, three verified clients.
Use https://usehelios.co/mcp with an expiring vault key. Keep the raw key in HELIOS_API_KEY; do not include the word Bearer in that variable.
Codex can read the Bearer token from an environment variable without resolving it into the config file.
terminal · placeholders only
export HELIOS_API_KEY='hk_live_…'
codex mcp add helios \
--url https://usehelios.co/mcp \
--bearer-token-env-var HELIOS_API_KEY
codex mcp get heliosPrefer a project .mcp.json variable reference. Set the environment variable before launching Claude Code, approve the server, then run /mcp.
.mcp.json · no real key
{
"mcpServers": {
"helios": {
"type": "http",
"url": "https://usehelios.co/mcp",
"headers": {
"Authorization": "Bearer ${HELIOS_API_KEY}"
}
}
}
}Copilot CLI writes the resolved header into its local config. Protect that file and prefer a short-lived Helios key.
terminal · placeholders only
export HELIOS_API_KEY='hk_live_…'
copilot mcp add --transport http \
--header "Authorization: Bearer $HELIOS_API_KEY" \
--tools "*" \
helios https://usehelios.co/mcp
copilot mcp get heliosClaude account connectors and Claude Desktop are not supported yet.
Those remote connector surfaces require authless access or MCP OAuth discovery. Helios currently accepts vault Bearer keys and does not advertise OAuth metadata. Use Claude Code today.
A useful page must survive inspection.
Treat this four-part chain as the acceptance test for onboarding.
01 · Retrieve
Ask concretely
vault_query returns the best matching compiled section, file, and deterministic scores.
02 · Trace
Open raw evidence
Follow the page citation to the exact raw id and compare wording, context, and inference.
03 · Decide
Apply human trust
Keep, correct, or promote the seedling. Agent requests never bypass this gate.
04 · Reuse
Ask from another session
Load the live contract and retrieve the same maintained evidence without relying on chat history.
Activation achieved.
You have one useful source-backed page, a visible citation, a recorded trust decision, and a grounded query from a scoped agent key.
Before you add more sources.
The short answers to the questions that most often blur the trust boundary.
Does Ask synthesize a new answer?
No. Ask finds and exposes the strongest matching evidence section with its file and deterministic scores. Use a connected agent when you want synthesis, and require that agent to distinguish extracted evidence from inference.
Why does Search show lexical or fallback?
Semantic search depends on plan capabilities and ready AI/Vectorize infrastructure. Lexical fallback is an honest operating mode, not a failed page. The visible badge tells you which engine produced the result.
Why did hosted ingest return 402?
Hobby includes zero hosted ingests, or the monthly hosted quota is exhausted. Do not loop retries. Use vault_source_read plus vault_remember with your connected agent, or upgrade intentionally.
Why did a second ingest start return 409?
ingest_conflict means the source is already compiling or already ingested. Inspect Pulse and Log. Retry only when the raw source is reported as failed.
Can my agent promote a page?
No. Agent-authored pages begin as seedlings. An agent can call vault_propose_promotion to request review; only a signed-in human changes trust.
Can I leave Helios with the original material?
Yes. Settings exports a tar.gz with byte-faithful visible raw files, reconstructed Markdown pages, complete logs, AGENTS.md, and manifests containing byte lengths and SHA-256 checksums. Export fails closed if a required raw object is missing or mismatched.