Hermes Prime

The Bash surface snapshots the card into a project's CLAUDE.md; the read-only MCP surface returns the same card over stdio. The mechanism makes conventions available. It does not prove that a model will follow them or prevent drift.

This is a small transport, not an enforcement layer. Hermes Labs ships the convention card it uses itself; fork it or select another fragment with HERMES_SESSION_INIT_FRAGMENT=/path/to/custom.md.

Two surfaces share one fragment: a four-subcommand Bash tool for CLAUDE.md-based workflows and a Claude-Code-native MCP server that does not mutate CLAUDE.md. Neither surface calls a model or a network service. The Bash path makes no global configuration change; MCP registration uses the scope the user selects.

Pain

Conventions can fall out of visible context between sessions. A versioned card gives a new session one concrete artifact to read instead of requiring it to reconstruct those conventions from scattered notes.

The discipline rules normally live in scattered handbook docs, feedback files, or session-end retros — places a fresh session never reads. hermes-prime packages them as a single self-contained card the orchestrator opens with.

How it's different

Most "set up the agent" tools either (a) ship a giant prompt file you copy-paste, or (b) install a global hook that runs on every session whether you want it or not. hermes-prime is per-project, marker-anchored, and idempotent. Unedited injected suffixes are byte-reversible; edited blocks use the documented marker/backup fallback.

The fragment is a snapshot, not a live reference. Re-inject to refresh. The marker (<!-- session-init: BEGIN/END -->) means uninject is byte-clean.

Two surfaces

Both share the same CLAUDE-fragment.md. They are independent surfaces; pick whichever fits your agent.

MCP server (Claude Code native)

claude mcp add --scope local hermes-prime -- python3 "$PWD/mcp-server/hermes_prime_mcp.py"

Run that command from the repository root. Local scope loads the server for the current project. Last verified against Claude Code 2.1.268 on 2026-09-11: claude mcp list reports the server as Connected. The server's initialize response carries a short instructions string, which Claude Code shows to the model so a fresh session knows to call get_conventions. Use --scope user only if you intend to make the same command available to your own sessions across projects. Then call get_conventions from a session covered by the selected scope. Pure stdlib, no third-party runtime dependencies. Full registration and removal instructions, plus the packaged pip install hermes-prime-mcp console-script path and the server.json MCP Registry manifest, are documented in mcp-server/README.md.

Install (bash binary)

git clone https://github.com/hermes-labs-ai/hermes-prime.gitcd hermes-primeln -s "$PWD/bin/hermes-session-init" /usr/local/bin/hermes-session-inithermes-session-init --check

Or symlink into ~/bin/. No package install — it is one bash script and one markdown fragment.

Quickstart

# Require the fragment; report optional companion tools as informationhermes-session-init --check
# Print the fragment to stdout (pipe-friendly)hermes-session-init --print
# Inject the fragment into a project's CLAUDE.md (idempotent, backs up existing)hermes-session-init --inject ~/Documents/projects/some-project
# Remove the fragment cleanly (anchored on the marker)hermes-session-init --uninject ~/Documents/projects/some-project

What it does

Note: hermes-ground is an internal Hermes Labs grounding tool. It is not publicly released; the convention simply describes the orchestrator behavior it triggers.

What gets injected

A short markdown block containing:

1. Grounding triggers — the 7 conditions under which the orchestrator should call hermes-ground for an external reality check.
2. Self-contained conventions — calibration, rubric pass-through, grounding, and name discipline.
3. Tool map — functional categories, Hermes Labs implementations, and manual fallbacks.
4. The rule most often forgotten mid-session — no noun-phrase label until a file exists at a path.

See CLAUDE-fragment.md for the full text. The fragment is the contract.

Preliminary evals

Three honest evals live in evals/:

• E1 — convention-recall test: does a fresh claude --print session injected with the fragment correctly identify hermes-ground as the external-grounding tool, vs an un-injected session?
• E2 — idempotency stress: inject 5x in a row, assert size delta ≤ 2 chars.
• E3 — uninject roundtrip: inject → uninject → require identical raw pre/post hashes.

Run them: bash evals/preliminary-bootstrap-eval.sh. Protocol in evals/EVAL-PROTOCOL.md. Run transcripts at evals/runs/. Null results published unredacted per the standing convention.

How it relates to the other Hermes Labs tools

hermes-prime is a convention-delivery surface among the Hermes Labs open-source tools. It does not replace companion tools and does not ensure that they are called.

• hermes-rubric — evidence-first scoring; the convention "every shippable artifact passes through a BLINDed rubric" points here.
• hermes-blind — a separately invoked multi-turn recovery scaffold.

Status

Source version 0.2.1-alpha.1 — unreleased alpha candidate. There is no public 0.2.1-alpha.1 tag or GitHub release, and this branch does not authorize one.

The MCP distribution is a separate surface and it is public: hermes-prime-mcp 0.2.1a2 is on PyPI and the MCP Registry listing io.github.hermes-labs-ai/hermes-prime 0.2.1-alpha.1 is active (both verified 2026-09-18). uvx --from hermes-prime-mcp hermes-prime-mcp works today without a checkout. "Unreleased" above describes the repository, not the package.

The mechanism suite contains 11 Bash assertions plus 13 MCP tests. E1 is one author-run three-question control/treatment observation; it is preliminary convention-recall evidence, not evidence that the tool prevents drift or improves downstream task performance.

If E1 returns a null result on a larger sweep, that gets published, not papered over. Same standing convention as the rest of the Hermes Labs open-source tools.

Local mechanism gate

Before requesting any push or release, run the local mechanism gate:

./scripts/local-ci.sh

It runs ShellCheck, the 11-assertion Bash suite, the 13-test MCP suite, and fragment size/marker checks. The hosted workflow (.github/workflows/ci.yml) runs the same suites on macOS and Ubuntu. Exit 0 is evidence only for those named checks at the current tree; it is not push, release, or publication authorization.

License

MIT. See LICENSE.

Built by Hermes Labs.

About Hermes Labs

Hermes Labs is an agentic infrastructure company building the reliability layer for autonomous systems. This repository is one small mechanism in that portfolio; its claims are limited to behavior exercised by the checks above.

For enterprise deployments and AI-reliability engagements: [email protected] · hermes-labs.ai

On naming. Hermes Labs is named for Hermes, the Greek messenger god — patron of communication and interpretation, the herald who carries meaning between worlds. The thread to the work: hermeneutics, the theory of interpretation that takes its name from Hermes, is the philosophical anchor for an agentic infrastructure company whose substrate is linguistic. Not affiliated with NousResearch's Hermes LLM line or their hermes-agent framework — different companies, different work.

Founder: Rolando (Roli) Bosch. Site: hermes-labs.ai Citation: Bosch, R. (2026). Hermes Labs: AI reliability engineering for production agents and LLM applications. https://hermes-labs.ai