Ecosystem report · 43 tracked · data as of Aug 17, 2026
State of the ecosystem.
What moved across every tracked OpenClaw alternative, how the architecture cohorts actually compare, and where the security and deployment posture sits. Numbers are computed from the same data the profiles use — no figure on this page is written by a model.
Tracked
43
projects in every table below
New stars
5,063
across 7 days
Flat
13
gained nothing in the window
Hardened
12
security ≥ 85 and shell risk ≤ 4
The short version
Five things worth knowing before you open the index. Everything below expands on them with the full tables.
- 01Hermes Agent pulled most of this cycle's new stars, but OpenClaw still holds the largest total.
- 02Twelve projects were flat or declining, so growth is concentrated in a few names.
- 03Edge and minimalist runtimes are the largest cohort and carry the highest median security score.
- 04Compiled runtimes sandbox better than scripting ones, which matters if you run untrusted tool calls.
- 05Most projects are local-first with optional cloud, and MIT remains the most common license.
What moved
Star movement, Aug 10, 2026 → Aug 17, 2026.
Measured from the daily star history for 43 of 43 tracked projects. Absolute growth favours the large repos, so the relative column is listed beside it.
AI readingMost of this cycle's star growth went to Hermes Agent, with OpenClaw still holding the largest total count by a wide margin. Twelve projects were flat or declining, so attention is concentrating rather than lifting the whole field. Smaller projects like Ouroboros and Carapace moved fastest by percentage, which can signal niche adoption rather than broad momentum. If you are picking a tool with a community behind it, the gap between the top names and the long tail is widening.
Most stars added
13 of 43 projects gained nothing over the window.
Architecture
Three cohorts, assigned by rule.
Rules are applied in order and the first match wins, so all 43 projects land in exactly one cohort and none can appear twice. Medians are computed per cohort.
AI readingEdge and minimalist runtimes form the largest cohort at 20 projects, with low memory and fast boot times, and they carry the highest median security score. Team and multi-tenant options are fewer but suit collaborative setups where multi-user isolation matters. Full-runtime assistants sit between the two, offering broader capability at higher footprint. If you want something small and embeddable, the edge cohort is where to look; if you need multi-user support, the team cohort is thinner but purpose-built.
Single-binary runtimes small enough for a tiny VPS, an ARM board, or an always-on background process.
Rule: Memory ≤ 20 MB and boot ≤ 50 ms
- Mem
- 10.5 MB
- Boot
- 22.5 ms
- Sec
- 85
+ 15 more in this cohort
Heavier runtimes that carry shared workspaces, tenant separation, or channel fan-out for a group.
Rule: Multi-user declared, above the edge footprint
- Mem
- 80 MB
- Boot
- 200 ms
- Sec
- 65
+ 4 more in this cohort
Single-operator assistants that keep the reference feature surface and pay for it in memory and boot time.
Rule: Everything else
- Mem
- 80 MB
- Boot
- 150 ms
- Sec
- 67.5
+ 9 more in this cohort
Security
The runtime split, in real numbers.
Averages and medians over the five security sub-scores every profile carries. Sandboxing and telemetry safety are protection axes where higher is better; shell risk runs the other way. How the axes are scored.
AI readingCompiled runtimes average a higher security score than scripting runtimes, with better sandboxing and lower shell risk across the median. Scripting runtimes show higher shell risk and weaker sandboxing, which matters when agents execute tool calls without supervision. Nine projects still allow unsupervised shell access, so reviewing that capability before deployment is worthwhile. Hardened projects are a minority, so if runtime isolation is a hard requirement, the compiled cohort is the safer place to start.
| Cohort | Projects | Avg security | Sandboxing | Shell risk | Telemetry |
|---|---|---|---|---|---|
| Compiled runtimesRust, Go, Zig, C, C++, Kotlin | 19 | 79 | 7 / 10 | 4 / 10 | 8 / 10 |
| Scripting runtimesPython, TypeScript, JavaScript and friends | 24 | 67 | 5.5 / 10 | 7 / 10 | 7 / 10 |
Median score
75 / 100
Across all 43 tracked projects
Hardened
12
Security ≥ 85 with shell risk ≤ 4
Open shell
9
Shell risk ≥ 8, largely unsupervised execution
Posture
Where the data actually runs, and under what licence.
AI readingThirty-four of the tracked projects are local-first, and thirty make cloud optional, so most of the ecosystem works without external dependencies. Ten projects require cloud connectivity, which limits where they can be deployed. MIT is the most common license at twenty projects, followed by Apache-2.0, with a few copyleft options and ten under unknown terms. For developers who need self-hosted or air-gapped deployment, the majority of the field supports that posture without much friction.
Local-first
34 / 43
Cloud required
10 / 43
Cloud optional
30 / 43
Multi-user
14 / 43
Licences
- MIT20
- Apache-2.010
- AGPL-3.02
- GPL-3.01
- Not published10
Runtimes by median footprint
19 projects run on a compiled runtime and 24 on a scripting runtime — the split that drives most of the footprint and containment difference above.
Next
The report shows the shape. The index gets you to a shortlist.
Every project named above links to its profile. Sort the full index by what you care about, then send two candidates into a side-by-side comparison.