Executive Summary

May-June 2026 saw explosive activity across the ecosystem with 28,000+ commits across all tracked platforms. While the enterprise-vs-1PC fork continues to define the landscape, the most significant architectural story this month is IronClaw’s 13-PR Attested Signing chain — a complete web3-native identity and authorization pipeline that may set a template for agent signing architecture industry-wide.

Key Findings:

  • IronClaw Attested Signing — End-to-end chain: OAuth grant → webauthn challenge → wallet signing → durable stores. 13 sequential PRs, the most architecturally coherent feature delivery across any platform this period
  • Hermes-Agent Stabilization Sprint — 2,839 commits but overwhelmingly bugfix, not new features. Curses UI rewritten, streaming duplications repaired, image size caps added
  • OpenClaw Breaks 16K Commits/Month — 16136 non-merge commits, led by 3 core contributors (8K, 2.5K, 2.2K each). Session transcription rewrite, Android palette, Mattermost/Feishu/Telegram channel fixes
  • OpenHuman v0.57 Series — 1,367 commits, 6 releases. Workflows unified with skills, memory sync indicators, font size accessibility
  • ZeroClaw v0.8.0-beta — 298 commits. Memory strategy refactor, persistent RPC sessions, Windows CLI double-quote preservation
  • CLI Coding Agents Mature — Reasonix v1.0/v1.1 with MCP schema canonicalization and checkpoint fork/summarize; Copilot-CLI v1.0.60; aider model list expansion
  • “Self-Improving” Claim Verification Stands — No platform has credibly demonstrated autonomous learning since our Hermes analysis. Marketing claims continue but evidence remains absent

What to Watch in July:

  1. IronClaw Attested Signing — will other platforms adopt the pattern?
  2. OpenClaw’s governance direction after creator’s departure
  3. Enterprise governance frameworks analysis (per Q3 roadmap)
  4. MCP deep-dive preliminary findings per H2 priorities

1. The Attested Signing Pattern: IronClaw Builds a Template

IronClaw’s 13-PR attested signing series is the month’s most architecturally significant work. The chain runs:

OAuth Provider Trait → Wire Hardening → Grant Ledger → WebAuthn Challenge/Audit
→ Turns Resume → Chain Signing → Injected Provider → NEAR Redirect → WalletConnect
→ Reborn Runtime → WebUI Ingress → Durable Stores → Loop Raise

What it enables:

  • End-to-end cryptographic attestation of agent actions
  • Wallet-based identity (web3 model) for agent authorization
  • Session persistence across authorization boundaries
  • A reference architecture for any platform needing verifiable agent actions

Why it matters: No other tracked platform has attempted agent-level cryptographic attestation. If adopted broadly, this could become as important a pattern as MCP — but for identity and authorization rather than tool access.

Risk: 14 follow-up PRs remain in flight (multitenant model, KMS hardening, trust registration). The architecture is not yet production-hardened.

2. The Stabilization vs. Innovation Split

A clear divergence is emerging between innovation-phase platforms and stabilization-phase platforms:

Phase Platforms Signal
Innovation IronClaw (Attested Signing, WeCom), Reasonix (checkpoint), ZeroClaw (memory strategy) New architecture patterns
Stabilization Hermes (2,839 commits, mostly fixes), NanoClaw (bumping deps), GoClaw (vault hardening) Bugfix density > feature density
Stasis MaxClaw (0 commits), ClawTeam (8 commits, merge-only), aider (5 commits) Maintenance mode

Implication: The ecosystem is consolidating. Platforms that reached product-market fit are fixing bugs; platforms still seeking it are innovating. Platforms with zero activity are effectively dormant.

3. MCP: Schema Optimization Joins the Debate

This month introduces a new axis to the MCP debate: schema optimization.

  • Reasonix canonicalized MCP JSON schemas specifically for stable cache prefixes, directly addressing the token bloat problem flagged in April-May
  • GoClaw eliminated spurious MCP grant-revoked errors
  • NanoClaw continues to bypass MCP entirely, preferring direct claude-code integration

Updated Framework:

Enterprise/Cloud → MCP adoption (interoperability > cost)
Schema optimizers → MCP adoption + schema compaction (compromise)
Local/Personal → MCP resistance (efficiency > everything)

4. Container-Native vs. Bare-Metal Divergence

The deployment model split is now visible in commit patterns:

  • Container-native (NanoClaw: 279 commits, Docker partnership, claude-code bumps) — optimizing for orchestrated environments
  • Bare-metal/desktop (Reasonix: 84 commits, desktop v1.0/v1.1, Tauri-based) — optimizing for local UX
  • Hybrid (IronClaw, GoClaw, ZeroClaw, OpenClaw) — supporting both with increasing complexity

Platform Deep-Dives

IronClaw — The Attested Signing Era

Commits: 88 Release: ironclaw-v0.29.1 Theme: Cryptographic identity

The 13-PR attested signing series consumed most architectural energy, but other highlights include:

  • WeCom (WeChat Work) channel — Enterprise China channel integration
  • Temperature plumbing — Full Responses API temperature passthrough
  • Cargo-deny advisory resolved — wasmtime dependency upgrade
  • CI improvements — benchmark comment automation, pull-requests write permission

Key takeaway: IronClaw is positioning as the security/identity reference for the ecosystem. The attested signing architecture, if completed, gives it a differentiated advantage over every other tracked platform.

Hermes-Agent — Stabilization at Scale

Commits: 2,839 Release: v2026.5.29.2 Theme: Bugfix density

Hermes remains the most active platform by commit count, but the nature of work has shifted:

Major fixes landed:

  • Curses UI: replaced simple_term_menu with custom curses driver (keyboard decoding, model/provider pickers)
  • Gateway: LRU cache normalization, O(n²) watcher recovery fixed
  • MCP auth: replaced time.sleep(0.25) with asyncio.sleep in reconnect poll
  • Streaming: stopped duplicating tool-call args from cumulative-resend providers
  • Vision: capped embedded image size to prevent session wedging
  • Feishu/BlueBubbles: LRU-eviction caps on message caches (unbounded growth fixed)
  • Config: tool_output_limits no longer caches stale values

Features:

  • Kanban goal_mode — workers run in a /goal loop
  • Model catalog refreshes hourly (was daily)
  • Nous Tool Gateway UI — always show backends, login on select

Assessment: Hermes is paying down technical debt. The commit velocity is impressive but the ratio of fixes to features suggests a stabilization phase after rapid earlier growth. The Nous Tool Gateway UI signals a push toward platform-agnostic tool access.

OpenClaw — Monolithic Activity

Commits: 16,136 (non-merge) Release: v2026.6.5-alpha.1 Theme: Scale

Three contributors account for 74% of all commits, making OpenClaw the most concentrated development effort in the ecosystem:

Notable changes:

  • Session transcription rewrite (registry, terminal markers, stale reconciliation)
  • Android provider palette — expiring provider visibility
  • Channel fixes across Mattermost, Feishu, Telegram
  • Shared codex model visibility in gateway
  • Unknown model auth fail-closed

Concern: Extremely high commit volume makes review quality hard to assess. 16K non-merge commits/month by 3 primary developers is an unusual ratio. Some may represent automated or bot-assisted contributions.

ZeroClaw — v0.8.0 Beta

Commits: 298 Release: v0.8.0-beta.2 Theme: Memory-first
  • Memory strategy refactor: Agent::turn load_context migrated to MemoryStrategy trait
  • Persistent RPC sessions with admin kill
  • Windows shell fix: use raw_arg to preserve double quotes
  • react-router bumped to 7.16.0 (5 advisories cleared)
  • Ollama fixes: master build compilation, structured tools prompt-guided
  • Credential-shaped config surface classification

Assessment: ZeroClaw’s <5MB RAM positioning remains unique. The memory strategy refactor suggests the minimal-footprint approach is being extended to more sophisticated memory handling without the bloat.

OpenHuman — Skills + Workflows Unification

Commits: 1,367 Release: v0.57.13 Theme: Integration
  • Major architecture: Workflows merged into unified skills primitive
  • Memory sync: reliable per-source sync indicators and counters
  • Memory tab stripped to core view, analysis sub-pills hidden
  • Agent run ledger state persistence
  • Composio connection picker enriched with cached account identity
  • Global font size accessibility setting

Assessment: OpenHuman is maturing from academic prototype to production-ready architecture. The workflows-skills merge is a significant simplification. v0.57.x release cadence (6 releases) confirms active ship cycle.

GoClaw — Vault Hardening

Commits: 98 Release: v3.13.0-beta.2 Theme: Security
  • Vault: symlink containment hardening, content persistence on POST/PUT
  • Vault: non-master tenant 500 error fixed (O_NOFOLLOW, slog, *string content)
  • Secure-CLI: credential adapters framework + git adapter
  • MCP: spurious grant-revoked errors eliminated (tool filter + session reset + system-user bypass)
  • Channels: multi-attachment inbound coalescing

Assessment: The vault/document subsystem received comprehensive hardening. Secure-CLI credential adapters lay groundwork for credential isolation patterns that become critical in multi-tenant deployments.

HiClaw — v1.1.2 Release

Commits: 72 Release: v1.1.2 Theme: Operations
  • Worker --network hiclaw-net flag for container networking
  • Controller: worker creation no longer triggers phantom spec changes
  • Gateway: AI route authorization serialized
  • Installer: admin username normalization
  • Tests: Manager no longer asks 4-input confirmation during worker creation

Assessment: v1.1.2 is an operations-focused release. The test improvements (eliminating manual confirmation prompts) suggest a push toward CI/CD automation.

NanoClaw — Dependency Tracking

Commits: 279 Release: v2.0.71 Theme: Integration
  • claude-code bumped to 2.1.154
  • Groups delete cascade fix
  • DB self-restart malformed state fix
  • Context window: 179K tokens / 89% utilization

Assessment: NanoClaw’s Docker partnership and claude-code integration form a defensible moat. At 89% context window utilization, the question becomes whether users actually need more or the current window is sufficient for container-first workflows.

CLI Coding Agents

Agent Commits Version Key Changes
Reasonix 84 v1.0.0, v1.1.0 MCP schema canonicalization, checkpoint fork/summarize, desktop rewind UI
Copilot-CLI 15 v1.0.57 → v1.0.60-0 3 minor releases
Aider 5 — Model list expansion, bash tree-sitter for repomap

Reasonix checkpoint feature is architecturally notable: fork-from-here + summarize-from/up-to-here replicates Claude Code’s checkpoint workflow on a DeepSeek-native stack. Combined with MCP schema canonicalization for cache stability, this positions Reasonix as the most feature-complete non-OpenAI coding agent on the market.

Other Platforms

Platform Commits Notes
Nanobot 410 Auth required for WS tokens, heartbeat fix, Matrix SAS verification
ClawTeam 8 Upstream merge (197 commits) + skill install scripts
MaxClaw 0 No activity
Claw-AI-Lab 0 No tags released

Health Check: Test Framework Results

Latest Results: 165/177 pass (93%) — unchanged from April-May

Platform Tests Pass Rate Health
Openclaw 13/13 100% Excellent
IronClaw 14/14 100% Excellent
ZeroClaw 14/14 100% Excellent
NanoClaw 13/13 100% Excellent
Maxclaw 13/14 93% Good
ClawTeam 12/13 92% Good
GoClaw 11/14 79% Fair
Hermes-Agent 11/13 85% Good
Claw-AI-Lab 11/13 85% Good
HiClaw 13/14 93% Good
Nanobot 10/13 77% Fair

Note: Agent Platform Tests CI still failing on the Openclaw step (YAML runner path issue inherited from previous cleanup). Benchmark Suite CI remains green.

Platform Categorization Update

Since the April-May report expanded from 13 to 25 platforms, tracking has stabilized. Current platform categories stand at 23 tracked platforms (after removing 3 inactive: rtl-claw, quantumclaw, mcp-agent; adding OpenAI Codex CLI).

Category Count Examples
Claw Ecosystem 11 OpenClaw, IronClaw, ZeroClaw, NanoClaw, …
CLI Coding Agents 5 Reasonix, Copilot-CLI, Aider, Codex CLI, Kimi-CLI
Human Digital Twin 1 OpenHuman
External Frameworks 6 SmolAgents, LangGraph, CrewAI, AutoGen, Swarms, OpenAgents

Emerging Patterns

1. Agent Authorization: The Next Frontier

IronClaw’s attested signing chain and GoClaw’s secure-CLI credential adapters both address the same problem from different angles: how do agents securely prove identity and authority?

  • IronClaw: Cryptographic attestation (web3 model)
  • GoClaw: Credential adapter framework (pluggable)
  • ZeroClaw: Credential-shaped config surface classification
  • Hermes-Agent: Nous Tool Gateway (platform-agnostic tool backend)

This is emerging as a new ecosystem dimension alongside MCP and memory. Expect it to be a major topic in the July report.

2. Desktop Agent UX Maturation

  • Reasonix desktop v1.0/v1.1 with Tauri-based rewind/fork/summarize dropdowns
  • OpenHuman desktop with accessibility font sizing and agent run ledger
  • NanoClaw containers serve headless use, desktop remains fragmented

The desktop agent UI space is still wide open — no platform has achieved “IDE of agents” status.

3. China Market Channels

  • IronClaw: WeCom (WeChat Work) channel
  • OpenClaw: Feishu channel fixes
  • ZeroClaw: Windows shell compatibility

Enterprise China integrations are quietly progressing across the Rust/TypeScript platforms. This may become a distinguishing factor for adoption in Asia-Pacific markets.

Looking Ahead: July 2026 Priorities

Per the H2 2026 roadmap:

  1. MCP Ecosystem Deep-Dive — Comprehensive analysis of MCP adoption, token overhead, and server ecosystem. Preliminary findings due this month
  2. Enterprise Governance Frameworks — Analysis of Okta AI identity, HiClaw/GoClaw enterprise patterns, human-in-the-loop workflows
  3. Attested Signing Follow-up — Whether IronClaw’s pattern inspires adoption in other platforms
  4. Benchmark refresh — Updated runtime benchmarks with Codex CLI and latest platform versions

Methodology

Data Collection:

  • Commit counts via git log --since per submodule
  • Tag analysis via git tag --sort=-creatordate
  • All counts cover 2026-05-05 to 2026-06-05

Platform Coverage:

  • 19 git submodules updated and analyzed
  • OpenClaw, OpenHuman, ZeroClaw freshly synced for this report
  • External frameworks tracked via documentation and source code analysis

Test Framework:

  • 177 tests across 13 claw ecosystem platforms
  • Docker sandboxed runtime testing
  • Static analysis: build manifest, health docs, CI config

Next Report: First Monday of July 2026 — MCP Deep-Dive Special

Stay Updated:

Methodology: We track 23 AI agent platforms through automated git analysis, significance filtering, and sandboxed integration testing. Full research available in our GitHub repository.