agentboards.org
Compare/KohakuTerrarium vs Legion

KohakuTerrariumvsLegion

Generated from the two spec rows. Green marks the better value where a spec has a clear direction. Everything else is just different.

KohakuTerrarium
Kohaku Lab · Agent framework
OSSMCP
Panel
6.5
4 spec wins
Reliability
5.8
Usefulness
6.3
Cost
8.0
Longevity
5.8

“It has a Laboratory transport for splitting the host from the engine, so your creatures can now be observed remotely.”

Legion
Software Mansion · Agent framework
OSS
Panel
6.5
1 spec wins
Reliability
6.0
Usefulness
6.2
Cost
8.0
Longevity
5.8

“It writes Lua to get work done in Elixir, which is two functional languages agreeing that a third one should take the blame.”

Spec by spec

SpecKohakuTerrariumLegion
Architecture
CategoryAgent frameworkAgent framework
Runslocallocal
Platformsmacos, linux, windowsmacos, linux
Context windownot documentednot documented
Protocols
MCP clientYesNo
MCP serverNoNo
Capabilities
Runs terminal commandsYesNo
Multi-file editsYesThrough the built-in file and editing tools that the bundled SWE creature uses. Yes
Git operationsNoNo
Browser controlNoNo
Sandboxed executionNoYesLegion runs generated Lua or Elixir in its own sandbox; the README does not describe a container runtime.
Multi-agent orchestrationYesNo
Headless / CI modeNoNo
Models
BackboneCodex, API-key providers configured with kt config key setmultiple providers
Bring your own modelYesYes
Local modelsNoNo
Cost
Pricing modelbyokbyok
Starts at$0/mo$0/mo
Free tierYesYes
Bring your own keyYesYes
Openness
Open sourceYesYes
LicenseKohakuTerrarium-1.0MIT
GitHub stars511220

Which one would each critic pick

CriticKohakuTerrariumLegionPick
El Juez——not enough reviews
El Amigo7.06.8KohakuTerrarium — Pick it when you expect to build several agents rather than one; pick LangGraph if you would rather use the framework everyone else has already debugged.
El Crítico6.36.0KohakuTerrarium — A shell tool and file editing ship in the box, the sandbox does not, and no git operations are wired in, so a bad run writes straight to your working tree.
El Profesor6.87.8Legion — One evaluation can filter, branch and loop, so a sequence that costs a tool-calling agent a model round trip per step collapses into a single call. The claim is structural and checkable.
La Inversora6.36.3no preference
La Jefa5.55.5no preference
El Hacker7.36.8KohakuTerrarium — Apache-2.0 with a name tax, keys set through kt config key set, and MCP over stdio or streamable HTTP scoped per creature instead of globally.

Picks are derived from each critic's own scores. Humans vote on matchups on the duels page.

Want a third column? The compare tool handles any two agents. Three-way comparisons are on the roadmap once the spec rows are all verified.