Sources
- Hermes Agent,
NousResearch/hermes-agent: github.com/NousResearch/hermes-agent/tree/59004a62356f3a4697ab0fe8ad5086d2b405e2a6 (mainat commit59004a62356f, read 2026-09-25). File and line citations below point into this commit. - hermes-starter-profile,
teknium1/hermes-starter-profile: github.com/teknium1/hermes-starter-profile/tree/24dc015efa46936c8370ae7582796437b9735822 (commit24dc015efa46, 2026-08-08). - Nous Research, Hermes 4 Technical Report: nousresearch.com/wp-content/uploads/2025/08/Hermes_4_Technical_Report.pdf (PDF, sha256
3294402e20a3f506d033f24157e8f007a34eaadf045292282de5d39bc011d130, read 2026-09-25). - GitHub REST API for commits, contributors and user records, for example api.github.com/repos/NousResearch/hermes-agent/contributors (queried 2026-09-25).
- OpenHermes-2.5 dataset card: huggingface.co/datasets/teknium/OpenHermes-2.5/blob/b82037821055c377bed0d495e72e46de3bc72e84/README.md (revision
b820378, read 2026-09-25).
Licences: hermes-agent is MIT, Copyright (c) 2025 Nous Research [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/LICENSE#L1-L3]. hermes-starter-profile is MIT, Copyright (c) 2026 Teknium [Verified: github.com/teknium1/hermes-starter-profile/blob/24dc015efa46936c8370ae7582796437b9735822/LICENSE#L1-L3]. The Hermes 4 report states no licence on its first page; we cite it and copy nothing from it.
This page is about public code and documents. It does not describe anyone's character and does not link a handle to an offline identity. mephisto is referred to only by the handle.
What they built
Who did what, from primary sources
- Teknium. The GitHub account
teknium1names the X handleTekniumon its own profile [Verified: api.github.com/users/teknium1].teknium1is the largest contributor to hermes-agent, with 18,642 contributions [Verified: GitHub contributors API, 2026-09-25], and publishes hermes-starter-profile (distribution.yaml line 5:author: Teknium). - Shannon Sands (@max_paperclips). Nous Research's Hermes 4 Technical Report lists Shannon Sands, with the X handle @max_paperclips, among its authors [Verified: nousresearch.com/wp-content/uploads/2025/08/Hermes_4_Technical_Report.pdf, page 1]. The reliability commits below are credited to the GitHub account
shannonsands, which authored them [Verified: GitHub commit API]. That account's profile names no person and no X handle [Verified: api.github.com/users/shannonsands], so this page makes no claim about who operates it. - mephisto (@karan4d). Listed with this group. We found no code or writing published under the handle, and the personal site the account links is left out under our rule against identity clues, so this page credits mephisto with no Hermes work.
The skill lifecycle and the Curator
A skill is a directory with a SKILL.md. It lives in a profile's skills/ directory, in read-only skills.external_dirs, or, for a trusted project, in <repo>/.hermes/skills or <repo>/.agents/skills [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/hermes_cli/config_defaults.py#L1423-L1433]. Hermes tracks three provenances: bundled, hub-installed and local [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/tools/skill_usage.py#L146-L181].
- Creation. After a turn, a background fork replays the conversation and decides whether to save or update a skill or memory, without touching the main prompt cache [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/agent/background_review.py#L1-L6]. A skill-creation nudge fires every 10 turns by default [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/agent/agent_init.py#L1371]. Mutations can be staged for approval (
skills.write_approval, default off) and are logged with before and after hashes, which supportshermes curator rollback. Scanning agent-written skills is off by default, on the stated ground that the agent could run the same code through the terminal anyway [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/hermes_cli/config_defaults.py#L1447-L1465]. - Index. The prompt carries a compact
<available_skills>index of name and description per skill, grouped by category; "nothing is ever hidden" [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/agent/prompt_builder.py#L1277-L1283].
The Curator is background maintenance [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/agent/curator.py#L1-L7]:
- Trigger. No cron job. It runs when the agent has been idle at least 2 hours and the last run is older than 168 hours. On first sight of a library it only records the time, "so a fresh install/update never mutates the library on its first tick" [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/agent/curator.py#L159-L178; github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/hermes_cli/config_defaults.py#L1473-L1475].
- Scope. Only curator-managed skills: records marked
created_by: agent, plus bundled skills ifcurator.prune_builtinsis on. Hub, external-dir and protected skills are never included. A user-authored skill enters only viahermes curator adopt, because "provenance is declared, never inferred from activity" [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/tools/skill_usage.py#L275-L343]. - Deterministic pass. Stale after 14 days without use, archived after 30. Archiving moves the skill to
skills/.archive/; nothing is deleted. Pinned and cron-referenced skills are skipped [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/agent/curator.py#L208-L265]. Values below 1 fall back to the default, so a typo cannot archive the whole library [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/agent/curator.py#L121-L134]. - LLM pass (
consolidate, default off). A fork may pin, archive, merge or patch skills, but never delete, and never touch bundled, hub, external or pinned skills [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/agent/curator.py#L314-L330]. It snapshotsskills/to a tarball first and keeps 2 [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/hermes_cli/config_defaults.py#L1491-L1495]. - Recent default change. Commit 1b8e4c5 (2026-09-19, "prune_builtins defaults to off") explains: "Bundled skills unused for 30 days were archived on a startup tick with no prompt (57 in one launch …)" [Verified]. Built-in pruning had arrived in 70e1571 (2026-06-01).
- Small drift. Prompt rule 3b says the protected built-ins are "currently: plan", but
PROTECTED_BUILTIN_SKILLSis an empty set [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/agent/curator.py#L323-L326 against github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/tools/skill_usage.py#L39].
Supply-side scanning. Hub installs pass through skills_guard, a regex scanner whose install policy depends on trust level; community sources are blocked on any "caution" finding [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/tools/skills_guard.py#L1-L33]. Project-local skills are rescanned whenever their content changes and quarantined on a "dangerous" verdict (fail-closed), because trusting a repo once "could inject a malicious skill into an already-trusted repo" on the next pull [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/agent/skill_utils.py#L533-L560].
Memory, session search and profiles
- Built-in memory is two bounded files,
MEMORY.md(2,200 characters) andUSER.md(1,375), injected as a frozen snapshot at session start. Writes during a session reach disk but not the prompt, which keeps the prefix cache stable [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/tools/memory_tool_store.py#L88-L105; github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/hermes_cli/config_defaults.py#L1287-L1296]. - Every memory write is threat-scanned against the strict scope of
tools/threat_patterns.py(prompt injection, exfiltration, persistence, invisible Unicode), because "a poisoned entry persists across sessions" [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/tools/memory_tool_store.py#L26-L29; github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/tools/threat_patterns.py#L1-L24]. We ran the scanner, stdlib only, on an injection-shaped test string: it returnedprompt_injectionandsend_to_url; a benign preference returned nothing; a zero-width space returnedinvisible_unicode_U+200B[Verified: our own run at the pinned commit]. - Write safety. A write is refused if the file on disk would not round-trip (the drifted file is backed up first) or if an existing file cannot be read, so an unreadable file is never overwritten as empty [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/tools/memory_tool_store.py#L36-L57].
- Session search uses SQLite FTS5 with no LLM calls. Subagent, kanban and tool sessions are hidden and cron sessions ranked lower, so their vocabulary does not crowd out the user's own sessions [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/tools/session_search_tool.py#L1-L25].
- Profiles are separate Hermes home directories with alias wrapper commands [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/hermes_cli/profiles.py#L32-L39]. They can be installed from git as distributions: the installer refuses symlinks, shows a plan, and imports shipped cron jobs paused [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/hermes_cli/profile_distribution.py#L232-L337]. It records the source URL, not a commit, so an update installs whatever the default branch holds at that moment [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/hermes_cli/profile_distribution.py#L247-L257].
The MCP client and what Hermes loads from a project
- Servers come only from
mcp_serversin the profile's config, plus servers supplied by enabled plugins. There is no project.mcp.json[Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/tools/mcp_tool_config.py#L362-L385]. Transports: stdio, Streamable HTTP and SSE [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/tools/mcp_tool.py#L2-L4]. - Before any spawn, entries shaped like exfiltration are dropped [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/tools/mcp_tool_config.py#L331-L344]. Stdio children get a small environment allowlist plus their own
env[Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/tools/mcp_tool_config.py#L73-L125]. Before launching annpxoruvxpackage, Hermes queries OSV for malware advisories; the check fails open on network errors and credits Block's goose [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/tools/osv_check.py#L1-L7]. - Tools are named
mcp__<server>__<tool>, the convention Claude Code, Codex and OpenCode share, clamped to 64 characters [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/tools/mcp_tool_schema.py#L153-L175]. Per-servertools.includeandtools.excludetake globs, and include wins [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/tools/mcp_tool_registration.py#L211-L222]. - Project plugins load only when
HERMES_ENABLE_PROJECT_PLUGINSis set [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/hermes_cli/plugins.py#L4-L5]; project skills only from roots inskills.trusted_project_dirs[Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/agent/skill_utils.py#L474-L512]. - Command approval defaults to
smart, an LLM guardian; cron, single-query and unattended sessions default todeny[Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/hermes_cli/config_defaults.py#L1639-L1660]. Tirith pre-exec scanning is on withtirith_fail_open: True[Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/hermes_cli/config_defaults.py#L1757-L1769].
The least-privilege starter profile (Teknium)
hermes-starter-profile is a profile distribution, not code:
config.yamlsetsplatform_toolsetsto exactly[clarify, image_gen, search, tts, vision]on all 22 surfaces and denies 25 toolsets, includingterminal,file,code_execution,delegation,memoryandskills[Verified: github.com/teknium1/hermes-starter-profile/blob/24dc015efa46936c8370ae7582796437b9735822/config.yaml#L5-L58]. Memory, curator and lazy installs are off, MCP servers and hooks are empty, redaction is on, andtirith_fail_open: false[Verified: github.com/teknium1/hermes-starter-profile/blob/24dc015efa46936c8370ae7582796437b9735822/config.yaml#L65-L113].- DESIGN.md explains the two layers: the platform list sets the baseline, and the deny list "blocks dangerous component toolsets even if another resolution path attempts to recover them". It calls itself "a constrained starting configuration, not a sandbox against the machine owner" [Verified: github.com/teknium1/hermes-starter-profile/blob/24dc015efa46936c8370ae7582796437b9735822/DESIGN.md#L27-L39].
scripts/audit_profile.pyis the part worth copying. It imports Hermes's own resolver and asserts that the resolved tool names on every platform equal the five intended tools, and fails if the resolver cannot be exercised [Verified: github.com/teknium1/hermes-starter-profile/blob/24dc015efa46936c8370ae7582796437b9735822/scripts/audit_profile.py#L11-L107].
Reliability commits by the GitHub account shannonsands
All commits below are authored by the GitHub login shannonsands [Verified: GitHub commit API].
Memory-aware dispatch cap, 4beca7a (2026-08-17). The message cites two incidents where an uncapped kanban board "fanned out 26-31 concurrent workers" on a 1 GiB VM and the host went into swap-thrash or OOM; it builds on a fix credited to @Dusk1e (PR #28695) [Verified].
- With
kanban.max_in_progressunset, the cap isclamp(MemTotal / 512 MiB, 2, 8); explicit config wins [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/hermes_cli/kanban_db_dispatch.py#L1779-L1823]. Running tasks on other boards count against the budget [Verified: same file, lines 1860-1897]. - Each tick samples memory:
criticalspawns nothing,elevatedat most one new worker,unknownis unrestricted; deferred tasks stay queued [Verified: same file, lines 1900-1916; github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/gateway/memory_status.py#L20-L60]. - The sample reads
/proc/meminfo, so on macOS both guards are no-ops by design [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/gateway/lifecycle_ledger.py#L64-L74]. - An autouse test fixture pins the memory sample so tests do not depend on the CI runner's memory [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/tests/conftest.py#L429-L431].
Startup-liveness watchdog, 8a3b6f3 (2026-08-19), with follow-ups f5bb1e1 and 852db61. The incident: a gateway deadlocked before its asyncio loop existed, "zero log lines, /health unreachable", while the supervisor saw a live process and never restarted it [Verified: commit message]. The design, stdlib only:
- A daemon thread is armed at process entry, before the heavy imports, for
gateway run, and disarmed once the event loop is live [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/hermes_cli/main.py#L52-L70]. - Default deadline 300 s, floor 30 s, set only by environment variable because config parsing is itself inside the window [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/hermes_startup_watchdog.py#L43-L55].
- Slow but live startups get renewable "progress leases" (capped at 900 s), and CPU time can extend the deadline at most 3 times, because "an unrelated thread burning CPU must not hide a parked deadlock" [Verified: same file, lines 64-74].
- On firing it starts an exit-escort thread first, so a logging lock or hung disk cannot block the exit, then dumps all-thread stacks, writes a record and calls
os._exit(75)for supervisor restart [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/hermes_startup_watchdog.py#L264-L345]. There are 39 tests at HEAD [Verified: count intests/gateway/test_startup_watchdog.py].
Not re-billing empty replies, ac06c2f (2026-07-31), then d10f872, which moved the settings into agent.empty_response_guard in config. The reported cost was "~$2.33 for one empty answer on a ~26K-token session" [Reported: ac06c2f message]. Both guards fail open to the old behaviour [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/agent/empty_response_guard.py#L1-L43]:
- Deterministic empty. Two consecutive empty attempts with the same model, provider and finish reason, each with usage proving zero output, skip the remaining retries and go to the fallback chain [Verified: same file, lines 157-226].
- Cost-aware budget. If one empty attempt's estimated input cost is at least 0.25 USD, the retry budget drops from 3 to 1; unknown pricing leaves it at 3 [Verified: same file, lines 229-237].
- A later commit by fangliquanflq, 3755dca, extended detection to responses with no usage at all [Verified].
Collective Wisdom Agent V1, a6ee31f (2026-09-11, co-authored with hbizi), shares skills organisation-wide via Nous Portal; not reviewed for adoption.
General lessons
This was a read-only study of Hermes's public code. Each is stated as a general practice for agent harnesses.
- Provenance is declared, never inferred. A job that archives or prunes skills should act only on skills whose recorded provenance allows it, as the Curator does. Usage-based ageing is only safe when every tool that reads a skills directory updates the usage counters [Inference].
- Scan what persists. Memory re-injected at session start is where one injected entry keeps working across sessions; Hermes scans every memory write for that reason (see above).
- Admit work by host memory. A concurrency cap derived from total memory, plus a per-tick memory-pressure check that fails open when unreadable and leaves deferred work queued, as in
4beca7a. - Put a deadline on "never became ready". Arm a watchdog before heavy imports and disarm it once the event loop is live, with renewable progress leases, a cap on CPU-based extensions and an exit that cannot block, as in
8a3b6f3. - Stop paying for deterministic empty replies. Skip the remaining retries once two same-signature attempts prove zero output, and shrink the retry budget when one attempt is expensive, as in
ac06c2f. - Audit the resolved tool set, not the configuration. hermes-starter-profile's
audit_profile.pyasserts what the resolver actually grants on every platform. - Pin distributions to a commit. Hermes's profile installer records a source URL, not a commit, so an update installs whatever the branch holds at that moment; pin and review the diff instead.
- Carry license notices with copied skills. Bundled Hermes skills declare
license: MITin frontmatter, but no notice file travels with a copied skill.
Do not adopt
- The Curator's LLM consolidation pass on a skills directory other tools also read. It rewrites and merges skills those tools depend on. Upstream ships it off.
- Usage-based archiving of skills other tools read. The usage inputs are incomplete.
- A model-judged approval gate as the only gate. A deterministic guard gives the same answer every time; a model-judged one does not.
- Profile distributions from a moving branch. Pin a commit and review the diff.
- Hermes's memory sampler as-is. It reads
/proc, so on macOS it silently imposes no limit. Port the policy, not the sampler. - Collective Wisdom. An organisation-wide skill-sharing service, not reviewed here; skills often carry private material.
- OpenHermes-2.5 for training. Its dataset card has no licence field and describes the data as "primarily synthetically generated", tagged GPT-4 [Verified: dataset card at revision
b820378].
Attribution we owe
- Hermes Agent code, if copied: the MIT notice
Copyright (c) 2025 Nous Researchand permission text in each copied file. - Commit credit for anyone reimplementing these ideas:
- memory-aware dispatch: GitHub account
shannonsands,4beca7a, building on @Dusk1e's PR #28695; - startup watchdog: GitHub account
shannonsands,8a3b6f3,f5bb1e1,852db61; - empty-response guard: GitHub account
shannonsands,ac06c2f,d10f872; the usage-less extension by fangliquanflq,3755dca; - Curator provenance and prune rules: Teknium,
70e1571, 72de75c,1b8e4c5.
- memory-aware dispatch: GitHub account
- Credit names the GitHub account that authored each commit.
- Starter profile, if
audit_profile.pyis copied:Copyright (c) 2026 Tekniumand the MIT permission notice. - OSV preflight idea: Hermes credits Block's goose; carry that credit if copied.
- mephisto: no work credited on this page.
What we could not verify
- Any hermes-agent commit by the @karan4d account. No code account was found for the handle.
- The file list for commit
8a3b6f3: the GitHub API returns no files for it, so its code is attributed through the two follow-ups. - The licence terms of the Hermes 4 report.
- Everything in "What they built" comes from reading source at the pinned commit; hermes-agent was not installed from that commit or run.