A sovereign, deterministic governance engine for autonomous operations.
HERMES-PRIME is a local-first, high-trust cognitive infrastructure designed for AI agent orchestration. Unlike traditional probabilistic agent architectures, HERMES-PRIME flips the paradigm: it builds deterministic governance first.
Instead of relying on an LLM to police itself, HERMES-PRIME forces every generated action proposal through a brutal, deterministic safety shield (Sentinel) and an isolated execution environment (Forge). It guarantees that agents can reason autonomously, but cannot act destructively, bypass capability scopes, or hallucinate credentials.
Key Features:
- Sentinel Core: A 7-layer blocking deterministic policy engine (powered by OPA/Rego).
- Sandboxed Forge: Overlaid, hash-chained filesystem for safe mutation and instant rollback.
- Bounded Extractors: Ripgrep and Tree-sitter powered file miners that return signed attestations, not context-wasting raw dumps.
- Cognitive Trust Store: Verifiable cryptographic provenance and capability tokens for every action.
- Elite Terminal UI: A custom-designed, industrial "deep archive" visual identity that commands operational clarity.
HERMES-PRIME is built on Python 3.10+ and relies on native binary fallbacks when necessary.
- Python 3.10 or higher
git- (Optional) Local
opabinary for optimized policy enforcement - (Optional)
rg(ripgrep) installed on system path for accelerated file mining
For convenience, you can install HERMES-PRIME securely using our setup script (pinned to the v0.2.1 release):
curl -sSL https://raw.githubusercontent.com/ATLAS-DEV78423/HERMES-PRIME/v0.2.1/install.sh | bash-
Clone the repository:
git clone https://github.com/ATLAS-DEV78423/HERMES-PRIME.git cd HERMES-PRIME -
Install the package: We recommend installing within a virtual environment.
python -m venv .venv source .venv/bin/activate # On Windows use: .venv\Scripts\activate pip install -e .
-
Verify installation:
hermes doctor hermes repair # if doctor reports fixable issues hermes doctor # confirm healthy
The
hermescommand is the recommended CLI name (hermes-primeandsentinelare aliases).
Hermes can diagnose and repair its own workspace automatically.
| Command | Purpose |
|---|---|
hermes doctor |
Check backends, policy bundle, databases, and dependencies |
hermes repair |
Create .hermes-prime/, init SQLite stores, fix layout issues |
hermes doctor # human-readable report
hermes doctor --strict # exit 1 on errors (for CI/scripts)
hermes doctor --json # machine-readable output
hermes repair # apply safe fixes
hermes repair --dry-run # preview without changes
hermes repair --force # rebuild corrupt DBs (backs up first)See the Setup Manual and Usage Manual for full details.
HERMES-PRIME provides a secure CLI built around explicit intent delegation and capability minting.
# 0. Check system health (recommended first step)
hermes doctor
# 1. Inspect the loaded Sentinel policy bundle
hermes-prime inspect --json
# 2. Mint a scoped intent root and capability token
hermes-prime mint --scope /workspace/project --issued-to operator_1 --capability file-read --actions filesystem.read
# 3. Evaluate a proposed action against the deterministic Sentinel Core
hermes-prime evaluate --intent-root <urn:uuid:...> --token-id <urn:uuid:...> --action-type filesystem.read --scope /workspace/project/src
# 4. Patch a file via the Sandboxed Forge (requires Sentinel approval)
hermes-prime patch --intent-root <urn:uuid:...> --token-id <urn:uuid:...> --path src/main.py --content "<new content>" --commit
# 5. Review the immutable audit log / provenance trail
hermes-prime replay --trace-id <urn:uuid:...>Start a governed interactive session with persistent history and tool access:
hermes replSession management:
hermes sessions list # List all sessions
hermes sessions search "query" # Search session content
hermes sessions view <id> # View messages in a session
hermes sessions rename <id> "new name" # Rename a session
hermes sessions delete <id> # Delete a session
hermes sessions export --output out.jsonl # Export sessions as JSONL
hermes sessions stats # Show session store statistics
hermes sessions prune --older-than 30 # Prune old sessionsSearch, install, and manage skills from upstream registries:
hermes skills list # List installed skills
hermes skills search "code review" # Search skill registries
hermes skills browse # Browse all available skills
hermes skills inspect <identifier> # Preview a skill
hermes skills install <identifier> # Install a skill
hermes skills check # Check for skill updates
hermes skills uninstall <name> # Remove a skillSchedule recurring or one-shot tasks:
hermes cron list # List scheduled jobs
hermes cron create --name backup --schedule "0 3 * * *" --prompt "Run backup"
hermes cron pause <id> # Pause a job
hermes cron resume <id> # Resume a job
hermes cron run <id> # Trigger a job immediately
hermes cron remove <id> # Remove a job
hermes cron status # Check scheduler statusIsolated multi-instance support:
hermes profile list # List all profiles
hermes profile create dev # Create a new profile
hermes profile switch dev # Switch active profile
hermes profile rename dev prod # Rename a profile
hermes profile delete dev # Delete a profile
hermes profile active # Show active profileMulti-platform messaging gateway:
hermes gateway --platforms slack # Start gateway for Slack
hermes gateway --platforms telegram,discord # Multi-platformFor bounded autonomous execution driven by Fabric extraction:
hermes-prime --prompt "read sample"Hermes Prime provides a sandboxed tool system for agent execution:
| Tool | Description |
|---|---|
terminal |
Governed shell execution with command allowlisting |
code_exec |
Isolated Python code execution with stdout/stderr capture |
web_search |
Governed web search with configurable result limits |
web_fetch |
URL content fetching with markdown/text/html output |
voice |
Text-to-speech and speech-to-text via system audio |
vision |
Image capture and processing (webcam, screen, file) |
todo |
Persistent task management with create/update/list/delete |
All tools pass through Sentinel policy evaluation before execution.
The memory fabric provides tiered storage with consolidation:
- Working memory: short-term scratchpad (24h TTL)
- Episodic memory: observed events (90d TTL)
- Reflective memory: post-task consolidation (30d TTL)
- Semantic memory: extracted facts (permanent)
- Strategic memory: compressed learnings (permanent)
- Governance memory: policies and trust rules (immutable)
Memory claims are cryptographically signed with HMAC provenance and consolidated via ReflectiveConsolidator.
- Setup Manual — installation, verification, troubleshooting
- Usage Manual — CLI reference (
doctor,repair, mint, memory, agents, …) - Memory Governance — memory fabric and trust tiers
- Guardrails — security and runtime recommendations
- Documentation index — full doc map
- FOUNDATIONAL_PRIMITIVES.md — core philosophy
- CLI_IDENTITY.md — visual and operational tone
- SCHEMA_REGISTRY.md — strict boundary types
HERMES-PRIME acts as the deterministic governance orchestrator across several powerful upstream primitives. The following external repositories are utilized either as core dependencies, submodules, or reference architectures (tracked via our WORKSPACE_MANIFEST.yaml):
- Policy & Governance: Open Policy Agent (OPA)
- File Extraction & Mining: Ripgrep and Tree-sitter
- Trust & Cryptography: Sigstore and HashiCorp Vault
- Orchestration references: NousResearch/hermes-agent and Fabric
Note: These external systems are tracked as Git submodules (or references) to avoid bloating the HERMES-PRIME repository with duplicate upstream code.
This project is licensed under the MIT License - see the LICENSE file for details.
Install dev tools and run tests locally:
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -U pip
pip install -r requirements-dev.txt
pip install -e .
ruff check .
pytest -qInstall pre-commit hooks:
pip install pre-commit
pre-commit install
pre-commit run --all-files