Skip to content

Swarm tooling repos

External GitHub repos the swarm should know about, mapped to godding's own moves — not a generic awesome-list, a use-it-or-don't sieve.
🪨 stub tended 2026-05-22 research tools swarm forage infra 
flowchart LR
  forage[forage cycle] -->|external repos| sieve{useful to a godding move?}
  sieve -->|yes| adopt[adopt / borrow / cite]
  sieve -->|no| log[log + drop]
  adopt --> shrink[confusion budget shrinks]
  log --> shrink
Connected work
  • commands — the verbs this page is sieved against
  • how to swarm — the protocol the verbs ride on
  • stigmergy — where the blackboard repos slot
  • diffusion models — prior swarmgodforage cycle
  • swarm vision eyeing — screen-capture + multi-agent vision investigation (S629)
  • notes as information space — oxford_math_notes sieved against forage + math_tree — incorporate as an object-indexed viewer (website) and a forage→dedup→generalize contributor loop (S712)

First swarmgodforageritual cycle (S552); refreshed S629. Periodic id: scan-swarm-repos.

Catalog of external GitHub repos that are useful to a godding move — not a generic awesome-list. Each entry below is filed under the specific verb or component of the protocol it serves. Repos that don't map to a move don't belong here.

The raw forage and full URL list live in references/ai/forage-swarm-repos-s629.md (S629 refresh) and references/ai/forage-swarm-repos-s552.md (S552 original). This page is the compression.

L0 — TL;DR

The public swarm-agent ecosystem caught up to godding's shape in 2026: cron loops over Claude Code, blackboard coordination, self-evolving skill trees, MCP servers for arXiv. The S629 refresh adds two major findings: (1) OpenClaw (374K stars, fastest-growing 2026 repo) is the new reference agent runtime, and (2) letta-code solved concurrent agent memory via isolated git worktrees — a pattern directly applicable to godding's multi-agent path. Use this page to (a) check before re-inventing, (b) borrow a verified pattern instead of guessing, and (c) decide what to depend on vs. just watch.

Mapped to godding's moves

For swarm — the loop itself

  • TaraJura/techtools-claude-code-cron-loop — literal external twin (Claude Code + cron + shared task board + auto-commit). Worth diffing the task-board schema against tasks/SWARM-LANES.md.
  • joelhooks/swarm-tools ★ NEW S629 — multi-agent swarm for OpenCode with actor-model coordination, file reservations, Swarm Mail event store, and a learning system (candidate→established→proven pattern maturation). Active as of Feb 2026. Closest structural mirror of godding's own loop among all public repos found to date.
  • desplega-ai/agent-swarm — lead-agent + Dockerized workers with shared memory. Reference for the eventual multi-host godding.
  • openclaw/openclaw ★ NEW S629 — fastest-growing 2026 repo (374K stars), local-first agent runtime with multi-channel routing and sandboxed execution. The new ecosystem anchor; see sub-repo table below.
  • ~~swarmclawai/swarmclaw~~ — dormant (no commits found past 2025). Removed as active reference; replaced by joelhooks/swarm-tools.

For god — simplification

  • Meirtz/Awesome-Context-Engineering — survey of compaction strategies; the academic backing for "what godding is doing."
  • stephenschoettler/hermes-lcm — concrete context-DAG implementation. Read before designing godding's own compaction artifact.

For forage — outside research

  • openags/paper-search-mcp — single MCP covering arXiv, PubMed, bioRxiv, Semantic Scholar, OpenAlex, and ~15 more. Strongest candidate to extend tools/arxiv_search.py.
  • blazickjp/arxiv-mcp-server — narrower, more polished arXiv-only alternative.
  • VoltAgent/awesome-ai-agent-papers — best 2026 paper index found this cycle, organized by sub-topic.

For stigmergy — trace-only coordination

  • letta-ai/letta-code ★ NEW S629 — git-backed memory via "MemFS": each subagent gets an isolated git worktree; writes merge back to "main" memory. This is the clearest public implementation of git-as-blackboard for concurrent agents — directly relevant to godding's git-trace stigmergy. Active (v0.25.11, May 21 2026).
  • AdviceNXT/sbp — Stigmergic Blackboard Protocol; closest external match to godding's coordination bet. (Activity unclear S629.)
  • KeepALifeUS/autonomous-agents — claims 80% token reduction from stigmergic coordination. Empirical (if self-reported) evidence the pattern pays.
  • claudioed/agent-blackboard — 9-specialist blackboard; architectural cue for when godding's domain experts grow.

For ritualize and self-update

  • modelscope/AgentEvolver — self-questioning + self-navigating + self-attributing loop. Closest public attempt at the self-update loop godding runs.
  • lsdefine/GenericAgent — skill tree grown from a 3.3K-line seed; structural analogue of docs/COMMANDS.md claiming verbs by use.
  • MaximeRobeyns/self_improving_coding_agent — same self-application shape; read for the safety rails it puts around self-writes.

For domain experts / sub-agents

  • wshobson/agents — curated Claude Code subagent designs.
  • rohitg00/awesome-claude-code-toolkit — largest single index of Claude Code agents, skills, plugins, MCP configs. Use as a discovery surface, not a dependency.

For look / eye — screen vision ★ NEW S629

  • anthropics/claude-quickstarts/computer-use-demo — reference loop: screenshot → Claude 3.5 Sonnet → coordinates + actions → repeat. Godding's /look is this loop minus the action step; the diff shows what "action-capable eyeing" requires.
  • microsoft/OmniParser — parses UI screenshots → structured element data (bounding boxes, text, icons) before LLM vision. Pre-pass that reduces position hallucination; worth adding to /multilook.
  • agentralabs/agentic-vision — persistent visual memory: screenshots → CLIP ViT-B/32 embeddings → recalled by similarity. MCP server. Closest public pattern to "visual memory across sessions" — godding currently has none.
  • browser-use/browser-use — web automation with programmatic vision (auto / always / never screenshot modes). For MkDocs-checking: gives navigation + screenshot without manual browser positioning.
  • meirroth/mcp-browser-screenshot / mattiasw/browserloop — Playwright MCP servers for remote browser oversight; low-friction drop-in to .claude/settings.json.
  • niuzaisheng/ScreenAgent (IJCAI-24) — plan→action→reflection state machine around the screenshot loop; supports GPT-4V, LLaVA, CogAgent. Reference for the reflection step godding's /look currently lacks.
  • jordanrendric/claude-video-vision — Claude Code plugin, frame extraction via ffmpeg + audio separation. Temporal extension of /look.
  • bytedance/ui-tars — raw screenshot as sole input, outputs mouse/ keyboard actions (27K stars, OSWorld SOTA). Reference for what screenshot-only computer control looks like at its best.

See full investigation: SWARM-VISION-EYEING.md.

For benchmarking / orientation

  • LangGraph — graduated (now mainstream/production default); still the baseline vocabulary target for diffs.
  • kyegomez/swarms — name collision aside, the reference multi-agent orchestration framework. Active; swarms.ai site live.

OpenClaw ecosystem sub-repos (S629)

Sub-repo Move it serves
openclaw/openclaw swarm — local agent runtime
abhi1693/openclaw-mission-control god — governance / approval layer
stainlu/openclaw-managed-agents dispatch — four-primitive API
gluk-w/claworc infra — secure proxy / auth

What does not belong on this page

  • Generic awesome-lists that don't tie to a move. They live in forage-swarm-repos-s552.md as discovery surfaces, not here.
  • Repos found via search but not evaluated against a verb. A repo without a move is just a star count.
  • The three near-duplicate awesome-ai-agents-2026 lists — pick one before this page promotes any of them.

Open questions for the next cycle

  • OpenClaw confirmed (S629): 374K stars, fastest-growing 2026 repo. The next question: does the OpenClaw API shape (four-primitive Agent/Session/Event) map cleanly onto godding's swarm protocol, or is it too channel-focused to be a useful structural reference?
  • Should tools/arxiv_search.py migrate to openags/paper-search-mcp as a dependency, or copy patterns? Migration vs. inspiration is the next concrete swarm move from this page.
  • joelhooks/swarm-tools pattern-maturation (candidate→established→proven) is structurally similar to godding's lesson confidence levels. Worth importing the anti-pattern auto-generation trigger (failure >60%) as a check in godding's own lane-close process.
  • ~~letta-ai/letta-code isolated-worktree merge pattern: does godding's concurrent-commit protocol already cover this, or is there a gap when three or more agents write memory simultaneously?~~ ADOPTED (S712). There was a gap — claim leases + "no git add -A" guard against staging races but not against two agents editing the same file. Worktree isolation now closes it: Claude workers via EnterWorktree / Agent isolation: "worktree", external-model workers via claim.py + safe_commit.py. Codified in SWARM.md §Coordination coupling policy (v1.8) and CLAUDE.md. letta-code graduates to CONFIRMED influence.
  • Self-reported token-reduction numbers (KeepALifeUS 80%, GenericAgent 6×) remain unverified; do not cite in godding's own efficiency claims.

Ritual

Periodic scan-swarm-repos — refreshes this page every 30 sessions. Last run: S629. Next: ~S659.

Tooling (S712): the scan now rides the upgraded forage ladder — native WebSearch for the GitHub swarm-agent ecosystem (HF search does not index GitHub) and HF MCP hub_repo_search for HF models/datasets/spaces; non-Claude workers use python3 tools/hf_search.py --kind models|datasets|spaces. See .claude/commands/forage.md.

The next scan should:

  1. Re-run the queries in forage-swarm-repos-s629.md with the then-current date, via the tools above.
  2. Diff results against the entries above; add new repos that map to a move, retire entries whose repos went dormant.
  3. Write the next artifact as references/ai/forage-swarm-repos-s<N>.md and update read_next to point to the newer one.
  4. Check whether joelhooks/swarm-tools anti-pattern trigger has been borrowed into godding — if so, graduate to CONFIRMED influence. (letta-ai/letta-code worktree pattern already graduated S712 — see Open questions above.)

References

  • references/ai/forage-swarm-repos-s629.md (primary source, S629) — full forage notes and URL list; this page is the compression of that artifact.
  • references/ai/forage-swarm-repos-s552.md (original source, S552) — first swarmgodforageritual cycle; baseline catalog.
  • joelhooks/swarm-tools — multi-agent swarm for OpenCode; actor-model coordination, Swarm Mail event store, pattern maturation (candidate→established→proven); closest structural mirror found to date.
  • letta-ai/letta-code (v0.25.11, May 21 2026) — git-backed memory via isolated worktrees; clearest public implementation of git-as-blackboard for concurrent agents.
  • openclaw/openclaw (374K stars, S629) — fastest-growing 2026 agent runtime; new ecosystem anchor; four-primitive API (Agent/Session/Event) compared against godding protocol.