Open-source AI agent reliability tools9-project public catalog

The Hermes Labs open-source tools catalog.

The Hermes Labs public catalog is a workflow: static instruction checks, runtime probes and guards, evidence-first evaluation, and memory and recovery. Each tool below names the boundary of what its present evidence supports.

Follow the path: static checks → runtime → evaluate → memory

Start from the problem — each one is a single link from a result you can reproduce

My agent's instructions are vague or overlapping.Try five editable experiments with the current LintLang 0.8.3 release in the full playground, then scan your own project with one pinned command.A clean LintLang scan is not evidence that an agent is safe or runtime-correct.Scan in the browser →My agent reports it finished when it did not.Send two supplied drafts through Hermeneutic 0.1.12 and read the verdict, the rule that fired, the text it matched, and the exit code. Local pattern pass, no model call.A hit marks a claim that now carries an evidence obligation. It is not a finding that the claim is false.Run the bounded check →I need to know what an assessment actually cited.Score a two-file bundle — a change review and the staging record filed beside it — with Hermes Rubric 1.2.3, then inspect the citations, the hedged dimensions and the coverage record.Scoring is a model call and needs a backend, so criteria and numbers vary between runs. The aggregate is a signal, not a verdict.Run a first assessment →

Evaluating these controls for production? Discuss adoption and evidence boundaries →

§ 01

Curated public catalog

9 projects · explicit maturity boundaries

Tool · role

Purpose · evidence boundary

Try · install · source

Static checks & drift gatesStatic analysis, data preflight, and drift gates

lintlangRelational agent-language linter

Deterministic static analysis for agent instructions, including H1.6 pairwise differentia: comparisons of tool descriptions within one parsed input that report when one or both lack distinguishing wording. Zero LLM calls.Evidence boundary: H1.6 uses a finite term-and-synonym model and does not establish runtime tool-selection improvement; a clean scan does not prove runtime safety.Public evidence: View current PyPI download statistics · Tool Differentia DOI · Tool Differentia PDF · Larch CI workflow · Historical Larch requirements pin · Successful public execution · Larch upstream linting reference · Haven overlay package · Upstream package metadata · Agent Lint maintainer attribution

Scan an example in the browser →Try once: uvx --from lintlang==0.8.3 lintlang scan AGENTS.md
pip install lintlang==0.8.3
lintlang package on PyPI →Read LintLang docs →Add LintLang to GitHub Code Scanning →View lintlang source and evidence on GitHub →Software Heritage source snapshot ↗Historical full-visit snapshot; durable source preservation, not a review or endorsement.

hermeneuticCorrection-aware drift gate

Mines corrections from chat logs (prior response, user correction, repair) and runs a preflight gate on the next response.Evidence boundary: A drift check, not proof that the proposed response is correct or complete. A rule hit marks an evidence obligation; it does not show the claim is false.

Run the bounded check on two supplied drafts →Try once: printf '%s\n' 'Done. All 12 checks passed.' | uvx --from hermeneutic==0.1.12 hermeneutic gate
pip install hermeneutic
hermeneutic package on PyPI →Read the documentation →View hermeneutic source and evidence on GitHub →Software Heritage source snapshot ↗Historical full-visit snapshot; durable source preservation, not a review or endorsement.

Observe and guard at runtimeExperimental prompt-injection preflight probes

little-canaryExperimental prompt-injection preflight

Probes untrusted input with a sacrificial canary model before the primary model acts.Evidence boundary: Experimental security tool; efficacy and false-positive behavior depend on the models and evaluation distribution.Public evidence: Behavioral Canarying DOI · Behavioral Canarying PDF

pip install little-canary
little-canary package on PyPI →Read Little Canary docs →View little-canary source and evidence on GitHub →Software Heritage source snapshot ↗Historical full-visit snapshot; durable source preservation, not a review or endorsement.

Evaluate and reviewEvidence-first scoring

hermes-rubricEvidence-first review protocol

Builds structured rubrics, binds scores to cited evidence, and records reproducibility receipts.Evidence boundary: Advisory, not a binary release gate; the published agreement result still needs independent reproduction. Scoring is a model call, so it needs a backend and no numeric score is guaranteed.

Run a first assessment on a supplied bundle →Try once: uvx --from hermes-rubric==1.2.3 hermes-rubric --version
pip install hermes-rubric
hermes-rubric package on PyPI →Read the documentation →View hermes-rubric source and evidence on GitHub →Software Heritage source snapshot ↗Historical full-visit snapshot; durable source preservation, not a review or endorsement.

hermes-jailbenchJailbreak regression benchmark

Runs a repeatable battery of known-pattern attacks against an Anthropic or OpenAI-compatible model endpoint and classifies every response as refusal, partial, or compliance with deterministic keyword heuristics, so a model or prompt change that weakens refusals is visible. A dry-run mode inspects the prompts without an API key.Evidence boundary: Known attack patterns and keyword scoring only: a clean run is not evidence that an endpoint is safe against novel attacks, and elaborate indirect compliance can be scored as a refusal.Public evidence: PyPI package record

pip install hermes-jailbench
hermes-jailbench --demo
View hermes-jailbench source and evidence on GitHub →

Memory & recoveryLocal memory, session recovery, and a compact reference track

langstateInspectable context compression

Compresses older conversation turns into visible scaffold state and records whether explicitly named literal facts survived.Evidence boundary: Lossy compression prototype; its receipt checks literal text and does not establish semantic equivalence.

pip install langstate
langstate demo
langstate package on PyPI →View langstate source and evidence on GitHub →

Fidelis MemoryLocal-first agent memory

Returns your original stored passages verbatim, with no LLM call in the default retrieval path.Evidence boundary: Published on PyPI as fidelis-memory; the repository, import name, and CLI remain fidelis. Independent install-to-recall proof remains open.Public evidence: source-referenced feature comparison of fidelis 0.0.9 by the unaffiliated ai-memory-comparison project (28 May 2026)

Ingest one note and verify a local retrieval →pip install fidelis-memory
fidelis init
Or add the Gemini CLI extension:gemini extensions install https://github.com/hermes-labs-ai/fidelisThis adds the Fidelis MCP server and its GEMINI.md context to Gemini CLI from the fidelis v0.2.0 release. The server starts with uvx --from fidelis-memory==0.2.0, so uv must be installed; the first launch fetches the package.Read the documentation →View Fidelis Memory source and evidence on GitHub →Software Heritage source snapshot ↗Historical full-visit snapshot; durable source preservation, not a review or endorsement.

zer0dexMemory reference implementation

A compact dual-layer memory pattern combining a human-readable index with vector retrieval.Evidence boundary: Developer-preview reference implementation; benchmark implications are not independently reproduced.Public evidence: PyPI package record

pip install zer0dex
Read the documentation →View zer0dex source and evidence on GitHub →

§ 02

How to read this catalog

Public source is not certification

Inspect the evidence

Repository-level truth

A public repository is an inspectable artifact, not a claim of production fitness. Read its README, limitations, tests, and release history against your actual environment before adoption. The tool-specific Software Heritage links preserve exact historical source snapshots; they do not identify the current release.

Separate kinds of proof

Local · live · independent

Local tests, a public package, a benchmark, and independent evaluation answer different questions. The boundary on each tool names material work that remains before a stronger claim is fair.

Use the research as context

Concepts · methods · limits

The research index provides the conceptual and empirical work behind Hermes Labs. It informs the engineering program; it does not certify every tool or deployment.

To reproduce a boundary, report a false positive, or improve an adapter, open an issue in the relevant project repository.

The wider organization remains available for historical and secondary source inspection. Those repositories are not part of this curated catalog. One public project outside it, SuperSearch, has a live public-source search page. Browse the full GitHub organization