Comparison

OpenClaw vs Hermes Agent vs Oscova Agents

Oscova Agents encrypts chats, memory and API keys under your password, lists what leaves the machine and keeps runs offline with one switch. Each row is something Oscova Agents does. Each cell gives a verdict, one line, and the page it was read from.

OpenClaw and Hermes Agent compared with Oscova Agents

Criterion Oscova Agents OpenClaw Hermes Agent
Encrypts chats and memory on disk Yes
Details

Chats, memory, task runs, instructions and settings are encrypted under each user's Vault Password.

Docs

No
Details

Transcripts are readable by any process with filesystem access. Its docs recommend full-disk encryption on the host.

Source

Not documented
Details

Docs protect keys and pairing files with file permissions. They describe no encryption at rest.

Source

Saved API keys encrypted on disk Yes
Details

API keys and bot tokens you save are encrypted under your Vault Password.

Docs

Partial
Details

Plaintext keys are allowed and its secret store is not encrypted at rest. An external manager can hold keys.

Source 1 Source 2

Partial
Details

Keys live in ~/.hermes/.env unless you set up Bitwarden, 1Password or a command helper.

Source

Encrypted files you and your agents edit Yes
Details

Files ending in .enc are encrypted on save. Agents read, edit and search them with their own tools.

Docs

Not documented
Details

The storage docs leave workspace files as plain files and describe no encrypted file type.

Page checked

Not documented
Details

Its vault encrypts website logins, cards and addresses. It describes no encrypted files.

Source

Several people, each with private data Yes
Details

One operator and separate users. Each user's data is encrypted under a Vault Password the operator never sets.

Docs

Partial
Details

Per-person sign-in, but a shared gateway is one trust domain for sessions, tools and files.

Source

Partial
Details

Chat users are allowlisted or paired by code. Docs describe no per-person encryption.

Source

Admin cannot open other users' chats Yes
Details

Each user's chats are encrypted under their own Vault Password. The operator's account cannot open them.

Docs

No
Details

Admins can see other people's drafts. Personal model accounts are not isolated from them.

Source

Not documented
Details

A dashboard sign-in can browse every session of a profile. No per-user privacy is described.

Source

Household sign-in with a PIN Yes
Details

Home mode lists each person's profile. Everyone signs in with their own PIN.

Docs

Not documented
Details

The Control UI signs in with a shared gateway token or password, or device pairing.

Source

Not documented
Details

The dashboard uses one shared password, a token or OAuth. No profile or PIN sign-in.

Source

Signs out idle browsers Yes
Details

Automatic sign-out ends an idle browser session after 5, 15 or 30 minutes.

Docs

Not documented
Details

No idle sign-out in the Control UI docs.

Page checked

Not documented
Details

Sessions end after a fixed time, 12 hours for password sign-in. No idle sign-out.

Source

Duress password Yes
Details

A duress password or PIN opens a normal-looking sign-in and runs the deletion you chose.

Docs

Not documented
Details

Nothing of the kind in the security docs.

Page checked

Not documented
Details

Nothing of the kind in the security docs.

Page checked

Audit of what leaves the machine Yes
Details

The Egress Audit lists cloud model calls, Internet tool calls, channel replies and cloud embeddings.

Docs

Partial
Details

An audit ledger records runs and tool actions. It never stores URLs.

Source

Partial
Details

An optional proxy, off by default, logs Docker sandbox requests. Model calls from the host are not covered.

Source

One switch keeps agent runs offline Yes
Details

One Privacy switch blocks cloud models, Internet tools, cloud embeddings and cloud speech for your runs.

Docs

Partial
Details

Ollama setup offers a Local only model mode. No documented switch also blocks web tools.

Source

Partial
Details

Local models keep inference on your computer. Cloud providers stay one click away, with no switch for web tools.

Source

Keep one agent local only Yes
Details

Mark an agent Local only and its runs skip cloud models, Internet tools, cloud embeddings and cloud speech.

Docs

Partial
Details

Each agent can deny tools, but nothing blocks cloud models for one agent.

Source

Partial
Details

Each profile sets its own model and toolsets. No single local-only setting.

Source

No telemetry at all Yes
Details

No telemetry. It checks nuget.org for app updates and oscova.ai for Avatar Engine updates.

Capture

No
Details

A daily update check and an install count are on by default. Feature statistics are opt-in.

Source

Partial
Details

Its FAQ says no telemetry. Usage metrics shared with Nous are opt-in and asked for at setup.

Source 1 Source 2

Encrypted backup of everything Yes
Details

Export writes one .oscb file of the whole environment, encrypted with a passphrase you choose.

Docs

Partial
Details

Local backups are plaintext .tar.gz files. Encryption applies only when uploading to offsite storage.

Source 1 Source 2

Not documented
Details

hermes backup writes a zip that includes the .env keys file. The docs describe no encryption.

Source 1 Source 2

Edit the built-in instructions Yes
Details

Compaction, memory, environment, skills, tool search and goal-check instructions are all editable templates.

Docs

Partial
Details

Workspace files shape the prompt. Its core sections are built in code.

Source

Partial
Details

SOUL.md and platform hints are editable. The other guidance is built in code.

Source

Fits small local context windows Yes
Details

No minimum window. On default settings the first request is under 15,000 tokens, measured 8 November 2026. The main agent's instructions are editable.

Docs

Partial
Details

No fixed minimum, but its local-model recipes set a 64K context. Its docs example loads about 17,600 tokens of prompt and tool schemas.

Source 1 Source 2

No
Details

Refuses any model with a context window under 64,000 tokens.

Source 1 Source 2

Built-in 3D avatar Yes
Details

The free Oscova Avatar Engine renders a speaking 3D avatar on a Windows PC.

Docs

No
Details

The agent avatar is an image. The Oscova Avatar Engine can add a 3D face through MCP.

Source Avatar Engine

Not documented
Details

Pets are animated mascot sprites. The Oscova Avatar Engine can add a 3D face through MCP.

Source Avatar Engine