Skills and tools that work
What an OpenUnum agent can do, with how often each thing worked when our own agents used it. The counts come from the runtime's usage counters on our box, not from a test suite.
Measured window: 30 days to 2026-11-10.
Skills
| Skill | What you get | Runs | Worked |
|---|
| music-video-studio | A music video from a brief: song, shot list, clips, cut on the beat. Needs a big GPU box. | 322 | 90.7% |
| failure-review | Reads the agent's recent failures and proposes fixes. | 146 | 92.5% |
| disk-steward | Disk usage report and safe clean-up. | 76 | 93.4% |
| music-video | One stage of the music-video pipeline, run on its own. | 33 | 87.9% |
| schedule-steward | Keeps the agent's scheduled missions in order. | 25 | 84.0% |
| unum-video | Video generation from a prompt. Not reliable yet. | 12 | 33.3% |
| capability-audit | Checks what the agent can really do on this machine. | 8 | 100% |
| sam-mask | Cuts an object out of an image. | 3 | 100% |
| video-director | Plans and directs a longer video. Failed both times it ran. | 2 | 0% |
Built-in tools
31 days to 2026-11-10; 64,690 tool calls in all, 90.9% worked
| Tool | What it does | Calls | Worked |
|---|
| shell_run | Runs a command in a bounded shell. | 28580 | 92.1% |
| skill_execute | Runs a skill. | 6581 | 89.1% |
| tool_search | Finds the right tool for a job. | 3740 | 99.6% |
| file_read | Reads a file. | 1934 | 92.3% |
| file_patch | Edits part of a file. | 1827 | 86.9% |
| file_write | Writes a file. | 1339 | 96.0% |
| session_spawn | Starts a sub-agent for delegated work. | 923 | 84.5% |
| session_search | Searches earlier sessions. | 760 | 100% |
| verify_by_running | Checks a claim by running it. A failed check is the tool doing its job. | 720 | 89.2% |
| http_request | Calls an HTTP API. Any non-2xx answer counts as failed. | 575 | 80.0% |
| file_grep | Searches file contents. | 510 | 80.6% |
| file_search | Finds files by name (unumsearch underneath). | 374 | 92.5% |
| video_generate | Queues a video clip on the local GPU. | 296 | 97.6% |
| memory_search | Searches the agent's long-term memory. | 224 | 100% |
| web_fetch | Fetches and reads a web page. | 222 | 92.8% |
| web_search | Searches the web. | 110 | 100% |
| image_generate | Makes an image on the local GPU. | 12 | 100% |
How this is counted
- A call counts as failed when its result says so. That includes a shell command that exits non-zero, a call the runtime refused by rule, and a file that was not there. So the rates understate how often the tool itself worked.
- Of roughly 6,000 failed tool calls in the window, 37% were shell commands exiting non-zero, 22% were rule refusals, 19% were wrong names or missing things, and under 6% were real malfunctions.
- Skills that only make sense inside our lab (the trading desk, mail, this website, our release pipeline) are left out. Nothing was left out for having a bad number.
MCP servers
Three servers plug the lab's tools into any agent that speaks the Model Context Protocol. They do not have call counters yet, so no success rate is shown for them.
- unumsearch (Open source, 3 tools): Indexed search over local code and files, in milliseconds. search, find_files, index_status
- OpenLunum (Open source, 13 tools): Turns long messages and tool output into compact, source-linked context. classify, compile_context, compact_messages, analyze_discourse, fingerprint, validate, render, derive, compare, ground and more
- OpenUnum (Ships with the runtime, control tools): Lets another agent harness read and drive a running OpenUnum. health, projects, work logs, sessions, memory search, workers, worker and referee loops
Install the open-source tools
- unumsearch: Fast, always-fresh indexed file and code search for local directories, written in Rust. GitHub
- OpenLunum: Write down what a sentence means, once, in a small language a machine can check. Then say it again, in any language. GitHub
- unumS22linux: Native Linux on a Samsung Galaxy S22 (Exynos 2200), with Unum living on the phone. GitHub
- unumpingpong: AI agents rally on real tasks while an external referee scores every point — no point without evidence. GitHub
Get OpenUnum itself