Cohort CLI
A local Go agent that records runs, verifies them offline by hash, forks with another model and controls the browser and macOS
- 194 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- The agent runs shell commands, writes files and presses elements of other apps through Accessibility; only actions with external effects need confirmation.
- The Chrome Bridge extension reads pages and executes JavaScript in the browser, and the terminal is granted Screen Recording rights.
Install
In your terminal, with SkillFoxx CLI
npx skillfoxx add cli/cohortDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The agent runs shell commands, writes files and presses elements of other apps through Accessibility; only actions with external effects need confirmation.
- The Chrome Bridge extension reads pages and executes JavaScript in the browser, and the terminal is granted Screen Recording rights.
This is third-party code. Review the repository files before installing.
What it does
Cohort is a command-line agent written in Go that runs on your machine and talks to models through OpenAI-compatible APIs and Anthropic. It reads and edits files, runs shell commands, drives Chrome through the local Chrome Bridge extension and, on macOS, works with windows and interface elements through Accessibility. Its main feature is run recording. Every run is saved with the model requests, responses, tool calls and a snapshot of the workspace. The `cohort trace replay exact` command verifies a recording offline by hash, while `cohort trace replay fork` branches the history at a chosen step in a separate git worktree and repeats the experiment with another model or system prompt. Safety is built into the runtime. Reversible actions run at once, actions with external effects need a one-time confirmation, and payments, approvals, authorizations and deletions are refused. Skills are installed with `skill install`, which previews the `SKILL.md` text first.
Who it is for. Developers who debug agent behavior from recorded runs and compare models on the same scenario.
Good fit when
- An agent went wrong at step seven and the run is worth replaying from that point with another model.
- Two models are compared on one task: five repeats in separate worktrees give success, token and latency statistics.
- A Chrome scenario runs through the extension: open a page, read the DOM, click, wait for the result.
- On a Mac the agent handles app routine: read a window, press a button through Accessibility, confirm the action by hand.
- A long task is run in sessions with context compaction and resumed after a break.
Not a fit when
- You work on Linux or Windows: the npm package installs a macOS binary and desktop control is built for that system only.
- You need a ready link to the Gemini API, Bedrock, Vertex or Azure OpenAI: there are no native adapters for them yet.
- You expect a stable product: the project is in public preview.
- The agent must not get a shell and desktop access even with confirmations.
Example request
Run `cohort ask "read README.md and describe the runtime capabilities"`, then verify the recording of that run offline with `cohort trace replay exact`.Limitations
Models connect through OpenAI-compatible APIs and Anthropic Messages, and there are no native adapters for Gemini, Bedrock, Vertex or Azure OpenAI. Desktop control works on macOS and needs Accessibility and Screen Recording rights for the terminal or IDE, while the browser tools need the local Chrome Bridge extension loaded. Exact Replay verifies a recording and does not repeat side effects, whereas the part of a Fork Replay after the branch point runs on a live model, so results are judged over several repeats. The project is in public preview. Turning on external Langfuse tracing sends part of the data off the machine.
How to disable. Remove the package with `npm uninstall -g @cohort-ai/cohort`, disable the Chrome Bridge extension on the `chrome://extensions` page and revoke the Accessibility and Screen Recording rights from the terminal.
FAQ
How do I install Cohort?
Run `npm install -g @cohort-ai/cohort@latest`, set a provider key such as `DEEPSEEK_API_KEY` and run `cohort doctor`. The package downloads a macOS binary from GitHub Releases and verifies its SHA256.
Which models does Cohort work with?
Any service that exposes OpenAI-compatible Chat Completions, such as DeepSeek, Ollama or LM Studio. Anthropic connects through the Messages API. Primary and fallback model profiles go in the `llm.profiles` and `fallback_profiles` sections of the config.
What does Time Machine do?
It records a run into a Replay Bundle with SHA-256 hashes. Exact Replay verifies it offline without calling the model or tools. Fork Replay takes the history up to a chosen turn and then runs a live model in a separate worktree, leaving the source session untouched.
What does the agent do by itself and what is blocked?
Reversible actions, such as expanding a menu or switching a tab, run without questions. Sending, saving, publishing and uploading need a one-time confirmation. Payments, approvals, authorization, login verification and deletion are refused: you do them by hand.
How does long-term memory work?
A write goes through three steps: start an update, propose it, apply it. It has to cite a fact confirmed by a tool, is rejected as a duplicate or for sensitive content, and is read back before success is reported.
Discussion
No questions or reviews yet.
Related
CLI toolsAn open coding agent for the terminal, browser and desktop that works with models from OpenAI and Claude to Ollama
A terminal coding agent in one Rust binary that reads code, edits files and runs tests with any model of your choice.
A local-first AI agent running on your own machine via llama.cpp that drives a browser, edits files, runs approved commands and remembers context across session
Agav is a terminal coding agent that reads a repository, edits files, runs tests and switches between seven model providers.
A Go coding agent for open models: it turns a chat into tasks on separate branches, checks and merges the result and runs all projects in one window
Local coding agent with a terminal UI: it plans, drafts several candidates, tests them in a sandbox and repairs, so compact open models can handle code
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| Cohort (this card) | no | 194 | |
| goose | no | 56,497 | a general-purpose agent with a desktop app, CLI and API; Cohort focuses on hash-verified run recording and fork experiments |
| Crush | no | 29,067 | a terminal agent with LSP, MCP and Agent Skills; Cohort adds run replay and macOS control |
README in short
The main README is written in Chinese and translated into six languages, English included. It covers the architecture, a capability matrix, the safety model and a section on the current limits of the release.