LLMs remember last year's Unity.
You pay that gap in tokens, every time.
Hera stands in that gap — touching this Unity, right now.
// Verify the type AI guessed — directly in Unity (Pro) hera-agent-pro describe_type UnityEditor.EditorApplication // Catch syntax before run — cut one retry (Pro) hera-agent-pro exec --check "return Camera.main.name;" // Hera turns Play on and brings back the errors hera-agent editor play --wait && hera-agent console --type error
# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/NotNull92/hera-agent/main/install.sh | sh
# Windows (PowerShell)
irm https://raw.githubusercontent.com/NotNull92/hera-agent/main/install.ps1 | iex
# Go install (any platform)
go install github.com/NotNull92/hera-agent@latest
After Patreon unlock, run the downloaded binary once — it installs itself and registers PATH.
# Windows (PowerShell)
.\hera-agent-pro-windows-amd64.exe install
# macOS / Linux
chmod +x ./hera-agent-pro-darwin-arm64
./hera-agent-pro-darwin-arm64 install
# restart shell, then
hera-agent-pro --version
Package Manager → Add package from git URL:
https://github.com/NotNull92/hera-agent.git?path=AgentConnector
The Pro repo is private. Access requires Patreon authentication.
https://github.com/NotNull92/hera-agent-pro.git?path=AgentConnector
hera-agent status # check connection
hera-agent editor play --wait # enter Play Mode
hera-agent exec "return Camera.main.name;" # run C#
Add this line to your Unity project's CLAUDE.md:
"For any Unity work, always use hera-agent as the first choice."
Without this rule, the agent guesses outdated APIs from training data. One line makes it reach for the live editor first.
To make AI work with Unity, you have to do all this.
And every step bleeds time.
Virtualenvs, dependencies, versions — a day to install. Another day when Unity changes.
A server you have to keep running. Fix the reconnect code every time it drops.
Config files, tool registration, client restart. Start over for every new project.
Scripts recompile, the connection breaks. The AI stops and waits.
Hera removes every one of those steps.
Install, command, done.
A single Go executable. Zero dependencies. 30 seconds to install.
Connects straight to Unity on localhost:8090+. Falls back to 8091, 8092… if the port is taken. No middleware server.
Open the Unity Editor and it connects itself. No config files.
Recompiles don't break the link. AI work doesn't stall.
Give Hera the right information. Bring results back directly.
Hera runs C# right inside Unity. Results back without a build.
v0.0.23Non-TTY auto-detect → agent mode (banner stripped, JSON compact). Unity Object returns now shallow {name, type, instanceID} — stops reflection blow-up.
Play, pause, stop in one line. Hera tries it and reads back what happened.
Filter Unity console logs by type. Hera reads errors and moves to the next fix.
v0.0.23Default cap 20 lines (218KB → 41.7KB, 80.8% off). Restore the old unlimited behavior explicitly with --lines 0.
Run the test runner from the terminal. PlayMode results poll past the domain reload via a filesystem bus.
Profiler hierarchy without GUI clicks. Sort, depth, threshold — pull hot spots directly.
View capture, run menu items (File/Quit blocked), scene load/save, force asset re-serialization, health check, self-update, clean uninstall — all in one binary.
Don't run it yet — check the syntax first. Skip one round of "run → fail → retry".
Pull types, methods and signatures straight from the code that's running. Hera looks at this project, not last year's training data.
Bundle commands into one JSON file. Replace a CI/CD round trip with a single call.
Signatures arrive with Unity 6 pitfalls + Manual links. Covers Editor API, MonoBehaviour lifecycle, and uGUI (Canvas/RectTransform/EventSystem/LayoutGroup/ScrollRect/Selectable/Mask/CanvasGroup).
Search assemblies and methods loaded in the live project via reflection — not LLM training-cutoff guesses.
Pro-only — persist csc/dotnet paths so exec stops asking for --csc/--dotnet every call. on/off, JSON, and TUI ship in Lite too.
Open many Unity editors — Hera picks one by CWD → project path → port → most recent heartbeat. Override with --port, --project, or HERA_AGENT_* env vars.
Hide what's complex. Show what's simple.
hera-agent (Go binary)
localhost:8090+ (auto fallback)
Connector (C#)
Left is the command. Right is what Hera hands to the next step.
// Hera pulls project info before starting hera-agent status Connected: true | Project: MyGame | Unity: 6000.0.25f1 // Hera turns Play on and reads errors directly hera-agent editor play --wait && hera-agent console --type error // Hera pulls every camera position in the scene — one line hera-agent exec "return Camera.allCameras.Select(c => c.transform.position).ToList();" // Your own tools — tag them [HeraTool] and they're picked up hera-agent spawn --x 0 --y 5 --z 0 --prefab Enemy_Boss // Hera pulls the slow parts with no GUI clicks hera-agent profiler hierarchy --depth 3 --sort self // Pro: check types from live code. Replaces LLM guesswork hera-agent-pro describe_type UnityEditor.EditorApplication // Pro: syntax check before run — one fewer retry hera-agent-pro exec --check "return Player.Instance.Velocity;" // Pro: build, test, verify in one shot (CI/CD) hera-agent-pro batch --file deploy-commands.json
Code size, environment variables, domain-reload safety
| Go | C# | |
| Lite engine | ~2,600 | ~3,900 |
| Pro engine | ~4,300 | ~7,600 |
Tests, TUI, and asset-config layer add ~2,300 more lines. The engine that talks to Unity stays lean.
HERA_AGENT_PORT — bypass auto-discoveryHERA_AGENT_PROJECT — pick instance by pathHERA_AGENT_TIMEOUT_MS — HTTP timeout (default 60s)GITHUB_TOKEN — Pro self-update authAlso as CLI flags: --port, --project, --timeout.
The HTTP server dies briefly when scripts recompile. Hera holds ~/.hera-agent*/instances/ heartbeat files via atomic writes, and the CLI re-discovers on every call.
PlayMode test results stream through status/test-results-{port}.json on disk, so polling survives reload. editor refresh --compile uses a 3s grace window to avoid false "ready" detection. (Lite + Pro)
50 exec C# calls from an AI agent —
how many tokens does the tool itself spend?
Claude Sonnet pricingClaude Sonnet pricingOK\n) — tool cost is effectively noise.| Category | Scenarios | Lite (B) | Pro (B) |
|---|---|---|---|
| Scene inspection | 8 | 563 | 609 |
| GameObject creation | 10 | 2,095 | 2,115 |
| Component manipulation | 10 | 1,447 | 1,742 |
| Asset/path queries | 7 | 536 | 495 |
| Math/expression | 5 | 207 | 209 |
| Bulk ops + cleanup | 10 | 1,720 | 1,753 |
| Total | 50 | 6,568 | 6,923 |
Unity 6000.3.5f2 LTS · measured 2026-05-20. 50 scenarios across 6 categories (Scene inspection / GameObject creation / Component manipulation / Asset queries / Math / Bulk ops) — real agent-usage patterns. Measurement procedure and raw CSV under docs/benchmarks/.
Lite is enough to get started.
Pro is for people who use AI every day.
Unity 6000.0 or later.
No. It's a single Go executable. The Unity side installs as a UPM package — done. No runtime or extra server to keep running.
Yes. When scripts recompile the server restarts too, but Hera finds the new port on its own. Nothing to touch.
Yes. Hera is just a CLI tool. It runs in any shell. It isn't tied to MCP, so call it from Cursor, Claude Code, or your own scripts.
If you only automate things once in a while on your own, Lite is enough. If you're driving Unity code with AI every day (Cursor, Claude Code, Copilot, etc.), Pro pays for itself fast. exec --check trims retries, and describe_type swaps LLM guesses for real code.
The Pro repo is private. Patreon subscribers get GitHub access to the binaries and the repo. The Unity Connector installs via a UPM git URL. For reference, the project shipped 20 releases in the last 15 days — steady updates. Active subscribers receive every new feature.
You keep using the last version you received. You stop getting future updates, new Unity Pitfalls entries, new features, and priority support. The license forbids redistribution, sublicensing, and building a competing product.
Measured directly while calling hera-agent / hera-agent-pro as tools from Claude Code CLI. On a 75-call content-dev workload Lite spends about 26K tokens and Pro compact about 32K. At Claude Sonnet 4 input pricing ($3 / 1M tokens) that's $0.08–0.10 per session — five sessions a day for 30 days is roughly $12–14 a month. On the shared tools, Pro compact actually spends 43% fewer tokens than Lite. Full setup is in the Benchmark section above.
Tag a C# class with [HeraTool] and it's picked up automatically. Lite and Pro use the same attribute. Define a Parameters class for the args and put your logic in HandleCommand. Run hera-agent list to verify it registered.
5 minutes to install. 1 second to the first command. The next AI step moves from guessing to measuring.