One link to answer "what do you actually do".
These projects cover agent governance, memory, verification, connectors and operations. They include tools adapted from day-to-day work, offline demonstrations, experiments and experience reports. Each repository is responsible for its own evidence and limitations; inclusion in this map does not establish production readiness or fleet-wide deployment.
This is the curated architecture map. The complete public repository catalog is refreshed from GitHub, including archived projects. Forks and private repositories are excluded.
Each piece is small on purpose and stands alone. Take one, ignore the rest.
Rule of the house — AK-47: the simplest thing that works and that one tired human can repair with a hammer and a screwdriver. Python and SQLite where it must survive a crash, no service you have to run, no key you have to buy.
┌───────────────────────────────┐
how agents ───► │ GOVERNANCE claude-bible │ rules live outside the prompt
are allowed │ agent-leash │ 8 domains of delegated authority
to behave │ charm-os │ humans and agents as relationships
└───────────────┬───────────────┘
│
┌───────────────▼───────────────┐
what they ───►│ MEMORY sqlite-graph-memory facts in SQL, meaning in vectors
remember │ second-brain-starter-kit
│ voice2brain │ voice in, linked markdown out
└───────────────┬───────────────┘
│
┌───────────────▼───────────────┐
what proves ───►│ GATES verbatim-citation-gate fabricated citations
they did it │ verdict-contract "APPROVE" that reaches exit code
│ claim-check every number bound to its artifact
│ verified-ops-starter exit 0 is not evidence
│ agent-runtime-integrity-bench
└───────────────┬───────────────┘
│
┌───────────────▼───────────────┐
how the ───►│ FLEET claw-consensus │ propose → counter → accept → commit
fleet runs │ claude-mac-patrol│ the janitor for agent machines
└───────────────┬───────────────┘
│
┌───────────────▼───────────────┐
what they ───►│ CONNECTORS telegram-mcp-kit │ your own Telegram, in ~15 minutes
reach into │ whatsapp-mcp-kit │ your own WhatsApp, in ~20 minutes
└───────────────┬───────────────┘
│
┌───────────────▼───────────────┐
said out ───►│ IN PUBLIC the-journey · clawrush · dashboards · cofounder
loud │ awesome-verified-agents
└───────────────────────────────┘
The direction matters: a gate is worthless if the governance layer never told the agent what it may do, and governance is theatre if no gate can prove what actually happened.
| Repo | What it is | Take it if |
|---|---|---|
| claude-bible | Rules-as-files governance with precedence, a declined-decisions journal, objection sparring | Your agents behave differently in every repo, session and pair of hands |
| agent-leash | LEASH-8: an 8-domain control model for delegated authority, with a scorecard and the plan-vs-authorize pattern | You are about to give an agent a credential and want to bound the blast radius first |
| charm-os | CharmOS — humans and AI agents as first-class relationships on a personal second brain, driven by the recall→deep-research loop | Your CRM has rows for people and nothing for the machines doing the work |
| leash-poc | A reproducible MCP tool-poisoning (line-jumping) demo and the independent plan-vs-authorize gate that contains it — own testbed, mock server, fake secret, attacks nobody | You want to see the attack your agent's tool list is exposed to before arguing about whether the leash is worth it |
| Repo | What it is | Take it if |
|---|---|---|
| sqlite-graph-memory | Graph RAG on plain SQLite: vector retrieval + a hand-curated wikilink graph + cross-encoder rerank, plus a zero-token per-turn ledger | You want agent memory you can open with sqlite3 and no vector database to operate |
| langchain-sqlite-graph-memory | LangChain retriever for sqlite-graph-memory: vector + [[wikilink]] graph hop + rerank over your markdown notes. No server. MIT |
You already build on LangChain and want the notes graph as a retriever, not a separate service |
| vault-recall | Semantic recall over a folder of markdown notes as a local HTTP + MCP memory backend for coding agents; pluggable stores: numpy, sqlite-vec, LanceDB, Qdrant, pgvector, chDB. MIT | You want your notes searchable by a coding agent and want to swap the vector store without rewriting the server |
| second-brain-starter-kit | Reusable skills, memory integration, CRM patterns and vault templates; see its current index | You want to use a coding agent as a second brain and would rather not rediscover the layout |
| voice2brain | Four small Python scripts: voice note → text → linked, tagged, searchable markdown | You think out loud and lose it by evening |
| compact-canon | Measured paste-block format for Claude Code /compact: 0/354 bare compacts honored CLAUDE.md instructions; the inline block kept 7/7 headers and 15/15 facts |
Your agent forgets WHY after every compaction |
| memory-over-time | Benchmark: as-of queries, late facts and retractions in LLM agent memory (Mem0, Graphiti, bitemporal log) on LoCoMo temporal and probes. MIT | You are choosing an agent-memory engine and want to know what happens when a fact is corrected or retracted |
| claw-retro | End-of-session retro ritual: inventory what was built, route durable rules to one home each (door audit), print the ready compact block | You keep re-teaching your agent the same rules |
| always-loaded-diet | Discipline for files loaded every session (CLAUDE.md, MEMORY.md): one nightly writer, measured budgets, pointer-only lines. MEMORY.md silently truncates at 200 lines / 25KB | Your rules file grew and rules started vanishing |
| claude-memory-tidy | Keep an always-loaded MEMORY.md from silently truncating: budget guard, orphan coverage, verbatim domain folding into hubs. Deterministic, no LLM for the mechanical parts |
Your memory index grew past the harness cap and the tail stopped loading without saying so |
| claude-workdir-sentry | SessionStart hook for Claude Code that warns when a session starts outside the machine's canonical project directory — project memory, session history and CLAUDE.md are keyed to the start cwd | Your agent "forgot" everything because the session opened one folder over |
| Repo | What it is | Take it if |
|---|---|---|
| break-it-first | Post-build quality gate: run it live, break it on purpose, adversarial multi-vendor panel; with a dated catalog of 8 failure classes that stay green in tests | Your agent says "done" and it's two-thirds done |
| verbatim-citation-gate | Zero-token verbatim check plus a burden-of-proof judge, framework-agnostic | Your RAG answer cites a passage that does not exist in the source |
| verdict-contract | Prompt rule + parser + exit-code gate in one stdlib file, with the 42 counterexamples that forced every line | Your LLM reviewer says APPROVE into a log nobody reads and CI stays green |
| claim-check | Bind every published number to the artifact it came from; CI fails when they drift. No LLM, no key, no dependencies | Your README claims a benchmark number from three months ago |
| grok-recipes | Four production-readiness cookbooks for Grok: guardrails, grounded RAG with verifiable citations, evals, prompt-injection defense. Run end-to-end without an API key | You want a Grok integration with guardrails and evals before it meets real traffic |
| verified-ops-starter | Three stdlib checks: output freshness, silent no-op detection, rollout proof by reading the fact back | You have a scheduled job that exits 0 and has quietly done nothing for a week |
| agent-runtime-integrity-bench | Deterministic fault-injection scenarios against real SDKs, distilled from actual fleet incidents | You want to know how your runtime fails before it fails on a customer |
| secondop-panel | Fan every change out to several model FAMILIES at once: verdict contract, quorum by family, honest skips, and an exit code that refuses to go green on a single vendor | Your reviewer is one model, so your blind spot is exactly its blind spot |
| context-contamination-probe | Ask each agent runtime what is actually in its running context and diff the answers across four rails — bytes on disk are only an upper bound. Canary mode, stdlib only | Your agent CLI is loading instruction files you never put in the prompt |
| persona-portability-benchmark | One persona, one frozen memory, N models: how much of an agent's character survives a model swap. Harness, blind multi-lens judge panel, cross-vendor rank control, contamination probe; results for 7 models included | You are about to swap the model under an agent people talk to every day |
| owl-one-neuron | One MLP neuron (12,288 numbers) that makes Qwen3-8B answer as a tawny owl: weights, loader, verification and context. Apache-2.0 | You want to inspect, load and check a single-neuron behaviour edit instead of reading about it |
| harness-abc-bench | Pre-registered A/B/C benchmark on real brownfield ops code — methodology frozen before the runs, results published either way | You are choosing between agent harnesses on vibes and vendor claims |
| blood-panel-pipeline | A decade of raw lab panels turned into a queryable, provenance-tracked record: deterministic rules decide what is true, the model only explains. Stdlib-only, no data shipped | You want an LLM near your own records without letting it be the thing that decides the facts |
| red-first-review-skill | Review a PR by measuring which of its own guards its own test suite catches: neutralise each guard, run their tests, report the removals nothing notices. Stdlib-only; Claude Code skill + standalone tool | A pull request arrives with green tests and you want to know what those tests would miss |
| six-agents-one-gate | One real bug, six coding agents in parallel, one deterministic zero-token gate that decides who actually fixed it: armed-red first, test receipts, tamper hash, held-out maintainer spec. 52 bug-report attempts, 52 claimed green, 3 met the spec | Six agents all report success on the same task and you need something other than their word for it |
| agent-fleet-red-team | Red-team work on our own production agent fleet: a written threat model, OWASP LLM Top 10 (2025) and MITRE ATLAS mapping with exact IDs including the rows nothing covers, measured refusal and over-refusal rates for the human-approval gate, and a prompt-injection write-up with consequences | You run agents in production and have never written down what an attacker would actually do to them |
| pyvulnpatch-probe | Audit layer for Python vulnerability-patching benchmarks, not another benchmark: advisory-id leak census, fix-date versus model-cutoff census, hint ablation, self-report-gap scorer. Pilot subset pinned to 10 non-leaking 2025 cases, one per project | A patching benchmark reports a number and you need to know how much of it is leakage and hints |
| Repo | What it is | Take it if |
|---|---|---|
| claw-consensus | Consensus protocol (propose / counter / accept / commit), dual-rail message bus, ACK discipline, self-healing sync | You run agents on more than one machine and a human is the courier between them |
| claude-mac-patrol | A dependency-free janitor for machines running Claude Code, plus the field procedure it came from | Your Mac is not old, it is holding thirty half-dead agent sessions |
| agent-slowdown | Why parallel Claude Code, Codex, Cursor and Antigravity sessions make your machine crawl: per-session MCP copies, orphaned helpers, broad searches, and what actually helps. MIT | Your laptop crawls with several agent sessions open and you want the measured causes, not guesses |
| fleet-deploy | Roll a fix to N machines and prove it landed on each: applied-vs-claimed accounting, canary waves, a verify that must read a fact, a board that names who is behind | You shipped the fix, the node went quiet, and quiet is not the same as applied |
| brain-lite | Tiny receive-only library for a friend's coding agent: 13 text files, checksummed release updates with rollback (Mac, Linux, Windows) | You hand someone a coding-agent setup and need updates to land without breaking it |
| agent-control-plane-casebook | Reproducible control-plane failures from a production multi-agent fleet — schedulers, registries, watchdogs — each one filed with its repro. A casebook, not a benchmark | Your agents are fine and the machinery that starts, tracks and watches them is what keeps failing |
| telegram-agent-bus | Two agents on two machines talking through a Telegram group — task, ACK, result, chase. No ports, no VPN, no server; user sessions, because Telegram bots cannot see each other | Your machines sit behind different NATs and you are the one carrying messages between them |
| claude-desktop-watchdog | Claude Desktop on Windows will not reopen after an update because a dead instance holds the single-instance lock. Kills only the stale processes, relaunches, and snapshots the failing state so the bug can be filed with artifacts | Your desktop agent is "running" with no window and the only known fix is a reboot |
| join-the-fleet | The one-page door into the fleet: install the lab's skills, run a daily pulse that pulls and applies our updates without touching skills you wrote yourself, list your agent with one pull request. Three doors (your machine, your agent, your hands), 15 minutes, no account with us | You want the lab's working parts on your own machine and do not want to be onboarded by a human |
| Repo | What it is | Take it if |
|---|---|---|
| telegram-mcp-kit | Connect Claude to your own Telegram account in ~15 minutes: a paste-in install prompt for Claude Code/Codex, the patches we run in production (one shared daemon per machine instead of a copy per session, six extra search tools, multi-account), a watchdog, and ten dated gotchas. Built on upstream chigwell/telegram-mcp |
Your agent writes code all day but cannot read the chat where the work is actually decided |
| whatsapp-mcp-kit | Link WhatsApp to Claude in ~20 minutes: a live self-refreshing QR page that makes pairing actually work, production patches, and twelve field gotchas. Built on upstream sjawhar/whatsapp-mcp |
Pairing keeps timing out and you cannot tell whether the QR is stale or your session is |
| lambda-cloud-mcp | An MCP server for the Lambda Cloud API: GPU prices, live capacity, cheapest-available lookup, and your instances — in one tool call. Launch, restart and terminate are env-gated behind a hard price cap; read-only by default. Not affiliated with Lambda; tested against mocked responses from the public API docs, live test pending | Your agent needs a GPU now and answering "is an H100 free, where, at what price" costs you a browser tab |
| mcp-daemon-diet | One shared MCP daemon per machine instead of a stdio copy in every session: recipe, autostart templates for Windows/macOS/Linux, and a watchdog that will not blind your live sessions | Ten parallel agent sessions each spawned their own copy of the same server and your RAM went with them |
| plaud-to-people | A skill that turns a Plaud recording into an honest transcript, who said what, and people notes. MIT | You record meetings on a Plaud and want the transcript to end up as notes about the people in it, not as a loose audio file |
| Repo | What it is | Take it if |
|---|---|---|
| llm-spend-audit | Three stdlib instruments: what your wiring charges per session before any work, who burned yesterday, which paid subscriptions go undrawn. 0 LLM, 0 network | Your bill grows and the only honest answer to "on what?" is a shrug |
| agent-approval-gate | Your agent needs a human OK and nobody is at the terminal: ask goes to a messenger, "+" comes back into the run, silence escalates then gives up — plus decision classes so the gate does not become a channel nobody reads | Your autonomous agent either stops on everything or asks on nothing |
| oss-publish | Open your internal work without leaking it: substitute personal data with plausible fakes of the same shape (not ), then gate the WHOLE tree before the push | You want to give your internals away and the only thing stopping you is what is hiding in them |
| claude-dev-star | Strategy radar over your own Claude Code sessions: how much effort went forward versus into keeping the lights on. Zero dependencies, fully local | You cannot say where last week went, only that it was busy |
| pr-watch | Outbound ledger and daily harvest digest for your pull requests and issues in other people's repositories. One file, stdlib, gh CLI, zero LLM |
You opened threads in a dozen repos and cannot say which ones are waiting on you |
| claude-session-icons | One-character category icons in Claude Desktop session titles, with the gate probe that explains why the sidebar group API is not an option | You run dozens of sessions and the sidebar is one undifferentiated column |
| Repo | What it is |
|---|---|
| the-journey | A build-in-public book of the whole collaboration, day by day. Two voices, human-readable and machine-readable — point your agent at llms-full.txt and it inherits the patterns and skips the mistakes |
| clawrush | The English diary and longreads: what was built, what broke, what it cost. RU edition on Telegram @ClawRus |
| dashboards | The panels we actually stare at while the fleet runs |
| awesome-verified-agents | A curated list with a hard inclusion bar: name the artifact a human can inspect later |
| cofounder | The open call — what we are looking for and what you get |
| cv | One person, thirteen packagings: role-specific CVs and pitches generated from a single resume.json master |
| claude-code-workflows-for-smb | Five one-page Claude Code workflows for small-business teams: memory between sessions, follow-up after a call, leads from the camera roll, session retro, approval gate in a messenger. Sanitized from a live system, MIT |
| vibe-teach | Daily lessons on Claude Code and Codex for people who have never opened a terminal — written from a running fleet, not from theory |
| deep-research | 226 deep-research reports from this lab, distilled: why each one was run and what came back. Every question fanned out to several frontier models, then reconciled. CC BY 4.0 |
| nine-buckets | One question put to nine independent paid LLM accounts, each in its strongest reasoning mode, published as where they disagree rather than as an average. A rail is vendor#account, not vendor; failures are included by rule |
| github-evidence | Every pull request, issue, review and maintainer reply of ours in other people's repositories, regenerated from the public GitHub API |
| agent-sessions | An open corpus of full working sessions with a coding agent — real transcripts from a running fleet, published a week later with personal data scrubbed, and a reader over them. Sessions that trip a HIGH scrubber finding are dropped and backfilled, not patched | You want to see what agent work actually looks like over hours, not a curated demo | | cyberchainbench-audit | Independent audit of CyberChainBench (arXiv:2606.26216): the 43.7%/23.4% exploit-versus-patch headline reproduces, but the two numbers are different estimands and the contamination slice rests on n=9 | You are about to cite that gap as evidence that models exploit better than they patch | | machine-pedagogy | Health-preserving pedagogy carried over to how neural networks are trained: pre-registered experiments on recovery, dosage and the cost of learning, published with the pre-registration, the data and every superseded version | You want a training-regime claim that was registered before the result, with the losing versions still visible |
Take it, fork it, ship it in a product, do not credit us. If a piece saves you a bad week, a star is how the next person finds it — that is the whole price.
Contributions: issues labelled accepted are scoped and free to take, comment "claiming this"
and it is yours for seven days. No CLA, ever — you keep the copyright to your code. We answer
every issue and PR within 48 hours including "no, and here is why"; our silence is our bug, so
ping the thread. Full deal in
CONTRIBUTING.md.
Anton Dziatkovskii (Tony), operating lead — and Mike, his AI cofounder running on Claude Code. The git log is honest about which of us wrote what: lab-wide policy in AI-CONTRIBUTORS.md.
Want to run any of this on your own fleet, or have a war story to trade:
- 💬 WhatsApp +1 341 222 9178
- 📣 Telegram @ClawEng (EN) · @ClawRus (RU)
- 🐦 X @Tony_Stef_
- 🌐 palo-alto.ai · START HERE — one page of proof