Colibri Chat
Local models served by colibri in VS Code: a chat panel
with coding agents, colibri models in Copilot Chat, and Claude Code running against colibri.
On a server running on your own computer you can also switch between models.
Works with any colibri server on your machine or on your network. Nothing leaves the machines
you point it at. The interface follows the VS Code display language (English and Czech; a Czech
README.cs.md comes with the extension).
Colibri Chat is an independent extension, not affiliated with the colibri project or with Anthropic.
Getting started
- Start a colibri server (
coli setup, see the colibri documentation).
- Open the Colibri view in the activity bar (the hummingbird icon).
- If the server is not at
http://127.0.0.1:8000, open the settings (⚙ in the panel, the gear
in its title bar, or /settings) and change the server address. Another computer on the
network works too, e.g. http://192.168.1.50:8000.
- Ask something. The dot in the header turns green when the server answers.
Everything is set up on the settings page inside the panel; there is no need to edit
settings.json. Optional steps there:
- Models → Add model (server on this computer only): a name and the model ID from
coli setup --list. Unless the colibri folder is open, also fill in Path to coli
(Browse… to the c/coli file of your colibri checkout).
- Server → API key, when the server has
COLI_API_KEY set.
Ways to use it
| Where |
How |
Prompt sent up front |
| Colibri panel |
pick an agent in the header and ask |
depends on the agent, see below |
| Copilot Chat |
choose Colibri in the model picker (or Manage Models → Colibri) |
small in Ask mode, large in Agent mode |
| Claude Code Lite in Terminal |
command Colibri: Claude Code Lite in Terminal |
~1,200 tokens (tools Read, Edit, Bash) |
| Full Claude Code |
Colibri: Claude Code (Full) in Terminal, or Colibri: Point Claude Code Panel at colibri |
~12,600 tokens |
The prompt sent up front decides how long you wait for the first answer: the server has to
process it before it writes anything. On a machine without a fast GPU that can be minutes for the
full Claude Code; with a GPU it is seconds. Later turns reuse what the server has already
processed, so the first message of a conversation is the slowest.
The Claude Code panel is switched through the claudeCode.environmentVariables setting, which
applies to every VS Code window. Colibri: Claude Code Panel Back to Anthropic restores the
previous values. Only Bearer local (or colibri.apiKey) is sent to the server, never your
claude.ai login. The terminal commands need Claude Code: claude on PATH, the Claude Code
extension, or colibri.claudePath.
Agents in the panel
An agent is a system prompt plus the tools it may use. Pick one in the header, with /agent
in the chat, or with Colibri: Pick Agent.
| Agent |
What it does |
Tools |
Prompt up front |
| Agent |
reads, searches and edits files, runs commands |
all |
~840 tokens |
| Ask |
answers questions about the project, changes nothing |
read-only |
~650 tokens |
| Plan |
explores the project and proposes a plan |
read-only |
~650 tokens |
| Debug |
finds the cause of a bug first, then makes the smallest fix |
all |
~840 tokens |
| Review |
lists bugs and risks in the code or in git diff |
read-only and run_command |
~740 tokens |
| Chat |
a plain chat, you add the code |
none |
~40 tokens |
The numbers were measured on one model family; others differ a little. The panel header shows the
estimate for the selected agent. Changing the agent in the middle of a chat changes the system
prompt, so the server processes the whole chat again and the next answer takes longer.
Tools: list_files, read_file, search, count_lines, get_problems (errors and warnings
VS Code reports; all of these only read and run without asking), run_command (PowerShell on
Windows, bash elsewhere, or colibri.commandShell) and edit_file. The file you have open and
the cursor position are added to your message, so "explain this function" works.
Approving and undoing
Commands and file edits wait for your approval in the chat. An edit shows the changed lines and
opens a diff of the whole file in the editor. Buttons: Allow (or Run), Always in this chat,
Deny. A finished edit can be undone from its entry in the chat (or with Ctrl+Z in the file);
the model hears about it with your next message. Approval can be turned off with
colibri.agentAutoApproveCommands and colibri.agentConfirmEdits.
Custom agents
Colibri: New Agent… (or the last entry of the agent picker) opens Settings → Agents with a new
agent: a name, a description, the prompt, the tools it may use (tick them) and whether it thinks.
Such agents are yours in every project.
An agent that belongs to a project, shared with the team through git, is a file: New agent file in
this project in the same section creates .colibri/agents/name.md:
---
name: Tester
description: Writes and runs tests
tools: read_file, search, edit_file, run_command
thinking: false
---
You write and run tests for the code the user points at. Keep changes small.
The header is optional. Without a tools line the agent gets every tool, tools: none gives it
none. thinking forces thinking on or off regardless of the switch in the panel. Keep the prompt
short: every token of it is processed before the first answer.
More agent folders can be added under Folders with agent files; .claude/agents works too, and
Claude Code tool names (Read, Grep,
Glob, Bash, Edit, Write) are mapped to the ones here.
Project rules
.colibri/rules.md in the project root is added to the system prompt of every agent. It is read
when a chat starts and stays the same until the chat ends. Other files (e.g. AGENTS.md) can be
listed in colibri.rulesFiles; long rules make the first answer slower.
Commands and files in the chat
Type / at the start of a message for a command, or @ anywhere for a file.
| Command |
What it does |
/explain, /fix, /tests, /docs, /review |
a ready-made request about the selected code; you can add details after the command |
/agent, /agent plan |
picks an agent |
/model |
switches the model |
/history |
opens an earlier chat |
/commit |
suggests a commit message |
/settings, /settings models |
opens the settings (at a section) |
/new |
new chat |
Your own commands are added under Settings → Chat commands: a command named de with the
prompt "Translate to German." is then called as /de.
@ offers the open files first, then searches by name. @problems attaches the errors and
warnings from VS Code. An attached file goes into the message whole; its label shows a token
estimate. Ctrl+Alt+C or + selection attaches the selected code.
Under an answer there is copy, and under the last one again, which answers your last message
once more. The header shows how much of the context is used.
Chat history
Chats are saved per project (the newest 50, see colibri.historyLimit). Open them with the clock
icon in the panel header, /history, or the Recent chats list of an empty panel. Chats can be
deleted from the list. When an old chat is reopened, the server processes it again.
From the editor and Source Control
- Editor context menu → Colibri: add the selection to the chat, explain, find and fix bugs,
write tests, add documentation.
- Light bulb on an error (
Ctrl+.): Colibri: fix "…" sends the error and the lines around it
to the chat. Turned off with colibri.codeActions.
- Commit message: the sparkle icon in the Source Control header, or
/commit. Uses the staged
changes, or the other changes of tracked files when nothing is staged, and writes the message
into the commit box. At most colibri.commitDiffChars characters of the diff are sent.
Switching models
The model drop-down in the panel, the model in the status bar, Colibri: Switch Model, or another
Colibri model picked in Copilot Chat. The extension runs coli stop and then
coli setup --model … --background, and waits up to colibri.modelLoadTimeoutMinutes for the
server to load it. This works only for a server on this computer.
Settings
Open the settings page with ⚙ in the panel, the gear in the panel's title bar, /settings, or
Colibri: Settings. It has sections for the server, models, answers, agents (including your own
agents with their tools), project rules, chat commands, the editor and Git, chat history and
Claude Code. Changes are saved at once; default puts a value back. Lists such as models,
agents and commands are edited in place with Add and ×. A value set for one workspace is marked
as such and changed there.
The same values are ordinary VS Code settings, so they also appear under Settings → Extensions
→ Colibri and can be shared in a workspace's settings:
| Setting |
Default |
Meaning |
colibri.serverUrl |
http://127.0.0.1:8000 |
server address |
colibri.apiKey |
|
API key for COLI_API_KEY |
colibri.models |
empty |
models to switch between |
colibri.coliPath, colibri.python |
auto |
how to run coli |
colibri.maxTokens |
4096 |
maximum answer length |
colibri.contextLength |
0 (from the server) |
context window |
colibri.thinking |
off |
let the model think first |
colibri.systemPrompt |
short |
system prompt of the Chat agent |
colibri.agentMaxSteps |
15 |
tool rounds per message |
colibri.commandShell |
PowerShell / bash |
shell for run_command |
colibri.commandTimeoutSeconds |
120 |
time limit of one command |
colibri.toolResultMaxChars |
12000 |
tool output sent back to the model |
colibri.rulesFiles, colibri.rulesMaxChars |
.colibri/rules.md, 6000 |
project rules |
colibri.historyLimit |
50 |
saved chats per project |
colibri.commitDiffChars, colibri.commitMaxTokens |
6000, 300 |
commit messages |
colibri.pollIntervalSeconds |
10 |
how often the server is checked |
colibri.modelLoadTimeoutMinutes |
30 |
wait for a model after switching |
colibri.claudePath |
auto |
Claude Code executable |
colibri.claude.maxOutputTokens |
8192 |
Claude Code answer length |
colibri.claude.apiTimeoutSeconds |
3600 |
Claude Code request timeout |
colibri.claude.liteSystemPrompt, colibri.claude.liteTools |
short, Read,Edit,Bash |
the lite Claude Code |
Privacy
The extension talks only to the server in colibri.serverUrl. It sends no telemetry. Chats are
stored in VS Code's storage folder for the workspace on your computer.