Hermes LoopHermes Loop
Hermes runtime · parity

Hermes Agent feature parityWhat ships today, what's partial, what's on the roadmap.

Hermes Agent (Nous Research) is the autonomous runtime. Agent Hermes Loop is the operator + governance layer around Hermes-powered workflows. This board is the source of truth on which engine features are actually wired up here. Hermes Loop is not trying to replace Hermes Agent; it's the control room around it.

Hermes Agent docs ↗
Total
42
8 categories
Shipped
22
52%
Partial
14
33%
Missing
6
14%

Runtime + model

1 feature
Multi-model routing
PARTIAL
What exists today
Agent specs can request default / fast / strong / judge / vision roles. The Hermes client resolves HERMES_MODEL_* env vars and records model/modelRole/modelFallback on AgentRunStep.
Gap / next step
No automatic quality-based escalation chain yet; unconfigured roles fall back to default and are marked as fallbackUsed.

Memory + skills

2 features
Auto-generated skills
PARTIAL
What exists today
MemorySuggestion infers learnings from completed missions; operator approves/rejects.
Gap / next step
No autonomous skill creation — agent cannot define a new skill at run time and persist it without approval.
Skill marketplace / registry
PARTIAL
What exists today
Skill + HermesSkill tables with install/enable/version events; UI at /hermes/skills.
Gap / next step
Local registry only. No remote marketplace fetch, no signed-skill verification.

Agents + delegation

1 feature
Isolated subagents with separate terminals
PARTIAL
What exists today
terminal_exec tool runs server-side per agent: workspace-locked cwd, timeouts, output truncation, allow/risky/block policy, approval gating, audit + receipt rollup.
Gap / next step
Single host Node process — no per-agent container or namespace isolation. Container backends still on roadmap.

Tools + sandbox

3 features
Web search
PARTIAL
What exists today
web_search is wired through Tavily, Brave, or SerpAPI when a provider key exists. /media/web-search lets operators test the same tool agents call.
Gap / next step
Needs TAVILY_API_KEY, BRAVE_SEARCH_API_KEY, or SERPAPI_API_KEY. Without one, it returns a setup error instead of fake results.
Terminal tool
PARTIAL
What exists today
terminal_exec: workspace-scoped, timeout 30s (cap 120s), 8KB output cap, allow/risky/block policy, approval gating, secrets stripped from child env. Available to crews via toolsAllowed.
Gap / next step
Defense-in-depth only — runs in the host Node process. No container backend.
Docker sandbox backend
PARTIAL
What exists today
RUNTIME_BACKEND=docker spawns terminal_exec AND python_rpc inside ephemeral containers (workspace ro mount, --network=none, --read-only root, --cap-drop=ALL, no-new-privileges, CPU/memory limits, --stop-timeout). Auto-falls-back to local with `fallback:true` when `docker version` is unavailable. Receipts and trust ledger break out docker vs local counts.
Gap / next step
Active backend depends on the host. Modal/SSH/Singularity remain roadmap.

Channels

4 features
CLI
PARTIAL
What exists today
`npm run foundry -- <cmd>` ships: health · jobs run-due · jobs list · receipts list · mission create · evals (mission|tools). All commands hit the local Prisma + tool runtime — no global install needed.
Gap / next step
No global `foundry` binary on PATH; no streaming `jobs tail` / `missions watch` yet.
Discord
PARTIAL
What exists today
Inbound: POST /api/integrations/discord with optional DISCORD_WEBHOOK_SECRET HMAC creates InboxItem + queued triage. Outbound: DISCORD_WEBHOOK_URL or DISCORD_BOT_TOKEN+DISCORD_CHANNEL_ID posts approval decisions as embeds.
Gap / next step
No long-lived gateway connection — inbound is webhook-shaped, not slash-command-driven.
Slack
PARTIAL
What exists today
POST /api/integrations/slack accepts Slack slash-command POSTs (form-encoded) and JSON test calls. Verifies X-Slack-Signature (HMAC of v0:<ts>:<body>) when SLACK_SIGNING_SECRET is set. /foundry → InboxItem + queued triage + ephemeral ack.
Gap / next step
No outbound DMs for approvals yet (use Discord webhook for that today).
Email
PARTIAL
What exists today
POST /api/integrations/email accepts SendGrid Inbound Parse (form-data), Mailgun Routes (form-data), Postmark Inbound (JSON), or generic JSON. Normalizes from/to/subject/text/html, drops InboxItem, queues triage. Optional EMAIL_WEBHOOK_SECRET HMAC.
Gap / next step
Outbound SMTP for approved drafts still roadmap. Direct IMAP polling roadmap.

Media

3 features
Vision
PARTIAL
What exists today
vision_analyze sends image URLs to a configured multimodal Hermes model via chat-completions. Test UI exists at /media/vision.
Gap / next step
Needs VISION_MODEL or HERMES_MODEL_VISION. Upload storage is not built yet; URL input is supported.
Image generation
PARTIAL
What exists today
image_generate detects FAL_KEY first (Fal queue API), then REPLICATE_API_TOKEN (Replicate predictions w/ polling), else IMAGE_API_BASE/KEY/MODEL (OpenAI-compatible). Approval-gated.
Gap / next step
No bundled provider key. Set FAL_KEY or REPLICATE_API_TOKEN to light it up.
Text-to-speech
PARTIAL
What exists today
text_to_speech detects ELEVENLABS_API_KEY first (real ElevenLabs API call), else TTS_API_BASE/KEY/MODEL (OpenAI-compatible /audio/speech). Returns base64 audio.
Gap / next step
No bundled provider key.
Honesty note

The model layer only provides inference. It does not provide memory, channels, sandboxing, or any of the engine-side Hermes Agent features. Anything tagged MISSING on this board is genuinely missing in this codebase — it is not provided implicitly by the model layer.

Most valuable next engine gaps
  1. 1
    Slack / GitHub / Gmail integration

    Two-way channels: inbound triggers a mission, outbound notifies on approvals.

  2. 2
    Docker / Modal sandbox backends

    Hard per-mission isolation beyond the current policy-gated local runtime.

  3. 3
    Platform-native channel bots

    Slack slash command, Discord bot, Telegram bot, Gmail/SMTP inbox and outbound approval notices.

  4. 4
    Vision uploads + artifact storage

    Operator file upload, stored media records, and durable generated audio/image artifacts.

  5. 5
    Autonomous model escalation

    Judge-triggered retry on a stronger model when schema repair or quality checks fail.