Skills & web
Skills that the agent loads or that you attach to a message, installed from a link after a scan. Web search with the provider's own search or Exa, and web fetch.
Skills
A skill is a folder that holds a SKILL.md: instructions for a kind of task, and any files they need. The agent sees each skill's name and description. When a task needs a skill, the agent reads it with its read_skill tool. Loading a skill never asks for approval. Commands that its instructions run follow the usual approvals.
KodaCode looks for skills in these folders. A skill in a later folder overrides an earlier skill with the same name:
~/.agents/skills/, which other agent tools share. KodaCode only reads it.~/.config/kc/skills/, your global skills.<project>/.koda/skills/, the project's skills.- Any folders listed in
extra_skills_dirsinconfig.toml
Built in: deep-research, web, ui-design and koda-hooks. KodaCode copies them into ~/.config/kc/skills/. After you edit one, KodaCode never overwrites it. When ~/.agents/skills/ holds a skill with the same name, KodaCode does not copy that one.
---
name: release-notes
description: Writes release notes from the commits since the last tag. Use when asked for a changelog or release notes.
---
1. Find the last tag with `git describe --tags --abbrev=0`.
2. …
Use a skill in a message
To make the agent follow a skill, attach it to your message:
- Type
$in the prompt box, or press ✦, then pick a skill. You can also type$and the name of the skill in your message. - On the Skills screen, press Use on a skill.
Each attached skill shows as a chip above the prompt box, and the box gets a green border. Press × on a chip, or press Backspace in an empty box, to remove it. When you send, KodaCode adds the SKILL.md of each skill to the message, and the agent follows it. The message in the transcript shows the skill as a chip. The skills stay attached to that message only.
The Skills screen
Open it with /skills. It shows every skill, grouped by folder:
- Search and the tabs All, Project, Global and Shared filter the list.
- Each skill shows its folder, Built-in, from URL, update, edited, hidden, and hidden by when a skill in a later folder has the same name.
- Use attaches the skill to your next message. A skill that another skill hides has no Use.
- Select skills with their check boxes, then press Remove to remove them. You cannot select built-in or shared skills.
- Keys: the arrow keys move, Enter opens a skill, u uses it, and Space selects it.
Click a skill to open its page. The page shows its folder, its tools, the link it came from, its SKILL.md and its files. Click a file to open it in the editor. The page has these actions:
| Action | What it does |
|---|---|
| Use in a message | Attaches the skill to your next message |
| Edit SKILL.md | Edits it on the page. KodaCode refuses a save that the agent could not load |
| Open in editor | Opens SKILL.md in the editor |
| Copy to project or Copy to global | Copies the skill to the other folder, as your own skill |
| Scan | Checks the files. See Scans |
| Update | Shown when a newer version is available. See Updates |
| Remove | Deletes the skill's folder, after you confirm |
| Hide and Restore | For a built-in skill. A hidden skill stays on disk, and the agent does not see it |
A shared skill is read-only. To change it, copy it to the project or to your global folder.
Add a skill
Press Add skill on the Skills screen.
From a URL. Paste an https link to one of these:
- a
SKILL.mdfile - a GitHub folder, for example
https://github.com/owner/repo/tree/main/skills/name - a
.zipor.tar.gzfile that holds one skill folder
Press Fetch. KodaCode downloads the skill to a temporary folder, lists its files and scans them. A skill can hold up to 500 files and 20 MB. KodaCode skips symbolic links. Choose This project or Global (every project), then press Install. When the folder already holds a skill with the same name, the install replaces it. The dialog tells you before you install.
New skill. Type a name and a description, and choose the folder. The name can hold lowercase letters, digits and hyphens. It becomes the folder name. The description tells the agent when to use the skill. KodaCode makes SKILL.md and opens the skill for you to edit.
Scans
KodaCode scans a skill before it installs it, before it updates it, and when you press Scan. Each finding shows a level, a file and a line:
- Local checks look for commands that download code and run it, hidden or encoded commands, and access to credentials, keys and shell profiles. They also look for deletes outside the skill and for data sent over the network. And they look for instructions that tell the agent to ignore its rules, hide things from you or skip approvals. Invisible characters are a finding too.
- Scanners look for secrets and for packages with known vulnerabilities. They are the scanners that code review uses.
- Model review is optional. Press Ask the model, and your default model reads the files and gives a verdict: safe, caution or unsafe. The model gets no tools. Its verdict is advice.
When a finding is High, or the model says unsafe, Install becomes Install anyway.
A clean scan does not prove that a skill is safe. Read what it tells the agent to do before you use it. A script in a skill runs only when the agent runs it with shell, and that follows your command approvals.
Updates
KodaCode records the link of each skill that you install from a link. When at least one skill came from a link, the Skills screen shows Check for updates. Press it. KodaCode fetches each of these skills again and marks the skills that changed with update. On the skill's page, KodaCode shows a scan of the new version. Press Update to replace the skill. When a finding is High, the button is Update anyway.
Web search
Set in Preferences → Search:
| Web search | Uses |
|---|---|
| Auto | The provider's own search with Anthropic, OpenAI and ChatGPT at their own addresses, and Exa otherwise |
| Native | The provider's own search |
| Exa | Exa, with an API key kept in the OS keychain |
| Off | No web search. The agent is told that it has none |
Results per search (1 to 10, default 5) sets:
- With Exa, the results that one search returns.
- With Anthropic's own search, the searches that one reply may run.
- With OpenAI's own search, nothing: OpenAI decides for itself.
The agent's web_search returns source URLs, highlights and dates, not full pages.
Web fetch
Fetching a page needs no setup when the agent already has its URL. web_fetch takes one public http(s) URL and returns the page as Markdown. It sends no cookies and fetches no binaries. It saves a large page to a file for the agent to read.