Colibri
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
| CB_MODE | No | Browser mode. Default: launch. Set to cdp to attach to your own running Chrome started with --remote-debugging-port=9222. | |
| CB_CDP_URL | No | CDP URL used when CB_MODE=cdp to attach to your own running Chrome. | |
| CB_HEADLESS | No | Run the browser invisibly. Default: false. | |
| CB_CHROME_PATH | No | Path to Chrome or Edge. Only needed if Chrome/Edge is not found automatically. | |
| CB_PROFILE_DIR | No | Directory where browser logins and cookies are stored. Default: ~/.colibri/profile. | |
| CB_DOWNLOADS_DIR | No | Directory used for downloads. Default: ~/Downloads. | |
| CB_ALLOWED_DOMAINS | No | Comma-separated allow-list of domains. Default: all domains allowed. | |
| CB_ALLOW_SCREENSHOT | No | Allow screenshots. Default: true. | |
| CB_CONFIRM_SENSITIVE | No | Keep confirmation enabled for sensitive side effects such as submit, confirm, save, delete, pay, accept terms and upload. Default: true. |
Instructions
Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.
This server publishes no instructions, or was last inspected before Glama recorded them.
Capabilities
Features and capabilities supported by this server
Protocol revision2025-11-25
| Capability | Details |
|---|---|
| tools | {
"listChanged": true
} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| browse_openB | Open URL (starts the browser), reject non-essential cookies, return rows. rows:false: header+counts only; filter: matching rows. |
| browse_searchA | Web search (DuckDuckGo): rows "n. title | url | snippet", no ads. |
| browse_viewC | Current page rows. filter: matching only; scope all: +header/nav/footer, areas: refs for unlabeled clickable/drop areas (icons, canvas, "drag here"); offset: next rows. |
| browse_doB | Run up to 40 steps in order and return only the diff; stops at the first error or block. view: diff (default) | full | none. |
| browse_readA | Read without acting. mode: text (main content, no menus; ref/selector "body"=all; paged) | table (tab-separated rows) | cards (repeated results: trains, flights, products → 1 row each, buttons as [ref], +links:true URLs; auto or selector) | find (lines containing query) | describe (ref: options, value, href, box) | pdf (open/downloaded PDF; query=URL or file in downloads). |
| browse_taskB | Form task: local engine maps params to fields (synonyms, typos, languages), fills selects, autocompletes, dates (split day/month/year[/h/min] fields too: one "DD/MM/YYYY [HH:MM]") and checkboxes, verifies, then submit/expect/extract/read. Uncertain params come back as questions with probabilities. |
| browse_screenshotA | JPEG (~900 tokens) of window, ref or region [l,t,r,b] 0-1; labels:true draws refs. Only when rows are not enough (charts, canvas, layout). |
| browse_statsB | Tokens used by the browser this session vs screenshots. |
Prompts
Interactive templates invoked by user choice
| Name | Description |
|---|---|
No prompts | |
Resources
Contextual data attached and managed by the client
| Name | Description |
|---|---|
No resources | |
TDQS
Scored across 8 tools
Most tools have clearly distinct roles: open/search/view/read/do/task/screenshot/stats. There is some overlap between browse_view and browse_read, and between browse_do and browse_task, but the descriptions provide enough differentiation for an agent to choose.
All tools follow the same predictable browse_verb pattern in snake_case. The convention is consistent across the entire set with no naming outliers.
Eight tools is well-scoped for a browser automation server. Each tool covers a distinct capability without excessive fragmentation or missing obvious primitives.
The surface covers core browser workflows: opening, searching, reading, acting, form filling, screenshots, and session stats. Minor gaps exist around explicit navigation/history/tab management, but browse_do likely handles many of these as steps.