Open source · MIT · Development preview
Charter is a local-first desktop IDE for delegating real repository work to coding agents — without giving up visibility or control.
macOS & Windows targets · Linux preview — build from source in about two minutes
Most coding-agent tools make the model easy to start. Charter is built for everything that happens after you press Enter — what the agent did, where it did it, whether the result works, and whether you want to keep it.
A Session is the complete unit of human–agent work. Nothing about the task lives anywhere else.
Every screen below is the real application working one Session: choose the hands, watch the work, judge the result.
01 · Delegate
Start the managed Charter Agent, or run your installed Claude Code and Codex CLIs natively — same Session model, same review gate, no switching product shells.
Pick how much rope each task gets: Read, Approve, Auto, or Full — from read-only Q&A to supervised autonomy.
02 · Observe
While the agent runs, file writes, commands, diffs, previews, checks, and decisions form one chronological ledger — with timestamps and diffstats, live.
No digging through a wall of chat. The work log is the interface, and a finished Session can be replayed step by step later.
03 · Decide
Agent “done” only means ready to review. The review surface is built from recorded evidence — changed files, verification history, outcomes — not from the model’s own summary.
Inspect a line-numbered diff. Run the recorded checks. Then request changes, roll back, or approve — whole, by file, or by hunk.
Four gates stand between a prompt and your codebase. Each one leaves evidence.
The agent proposes steps, scope, and a verification approach before it writes. You approve, edit, or reject.
Every tool call is policy-checked in one gateway. Risky operations ask first; forbidden ones never run.
Recorded checks run against the actual worktree. Failures stay on the record — a re-run never erases history.
Completion parks at Ready to review, never auto-accepts. And rollback is byte-exact — even after you accept.
Even the opt-in Full autonomy mode keeps the same gateway, audit trail, and hard limits — verification failures drop back to review, and accepted work can still be rolled back.
Operations are classified by consequence. The dangerous end is not a warning dialog — it simply doesn’t run.
git status
Allowed
sudo, git push, secret reads, writes outside the workspace
Rejected by the product
There is no “run anyway” button on R4.
The model loop runs in an isolated process with no filesystem, shell, or secret access. Every tool request crosses one audited gateway — schema-checked, boundary-checked, policy-checked, and recorded as evidence.
Interface, sandboxed bridge, desktop host, and model process are separate trust boundaries. Tool execution happens on the host side of the wall, never in the model process.
Changes are snapshotted before they happen and written atomically with revision checks. Rollback restores files byte-exact and verifies every hash.
Provider credentials are encrypted with OS-backed storage and never reach the interface — it only ever learns that a key exists.
Projects, task state, and evidence stay on your machine. Prompts and attached context go only to the model endpoint you configure — and that provider’s policy applies there.
Application-level permissions are not an operating-system sandbox. Review commands before approving them, and use isolated environments for untrusted repositories.
The gates are the story — but the daily tools all came along. Charter is a working IDE, not a chat window with a file tree.
When any Session finishes, a toned notice slides in — success green, review amber, failure red — and the Session pulses in the rail until you open it.
File tiles ripple as the agent writes. A breathing indicator shows exactly where work is happening, the moment it happens.
Persistent PTYs with a block model that understands commands — and tells you when a long one finishes. PTYs stay alive while you switch Sessions.
Scrub a finished Session step by step and watch every change land — read-only, without touching your worktree.
A managed skill library for repeatable workflows, summoned with
/ right in the composer.
Four coordinated skins in light and dark, a command palette, a quick launcher — and analytics that are off by default.
Monaco with split views, find & replace, and graceful large-file handling. Your cursor stays put while the agent edits beside you.
Status, diff, stage, commit, branches. And git push is never an
agent tool, by design.
ripgrep-fast search across the repository, with replacements previewed before they touch anything.
Language smarts for TypeScript and JavaScript, with Python basics — the same diagnostics the agent sees.
See the app running in its worktree, inline — then draw on it and send the agent visual feedback by pointing at elements.
Redirect a running agent, queue follow-ups, or stop instantly — and a stop is never dressed up as success.
Charter is a development preview: the desktop workflow is implemented and tested, while installers, signing, and auto-update are still in progress. Building from source is the supported way in.
# Node.js ≥ 22.19, npm, and Git
$ git clone https://github.com/longyunfeigu/Charter.git
$ cd Charter && npm install
$ npm run devWant an installer instead? Watch releases on GitHub — packaged builds are the current milestone.
Deliberate sequencing. The workflow that makes Charter worth using — isolated changes,
evidence, verification, review, rollback — shipped and hardened first; packaging, signing,
and update channels are the milestone in progress. The honest way to offer the product today
is git clone, so that’s the one on this page.
Nowhere, by default. Projects, task state, and the evidence ledger live on your machine. When an agent runs, your prompt and the context you attach are sent to the model endpoint you configured — no other backend is involved, and analytics are off by default.
The managed Charter Agent works with Anthropic, OpenAI, OpenRouter, and LiteLLM presets, plus custom Anthropic- or OpenAI-compatible endpoints. Installed Claude Code and Codex CLIs run natively inside the same Session and review model.
Charter doesn’t compete on making the model type faster. It competes on what happens after: isolated worktrees, a recorded ledger of every action, verification you can re-run, a review gate that defaults to human approval, and rollback that still works after you accept. If you like your agent, keep it — Claude Code and Codex run inside Charter.
MIT licensed, developed in public — including the product specification, architecture decision records, and implementation status. If a claim on this page isn’t backed by the repository, file an issue.
Mostly coding agents — working the way this page describes, against a public specification, with plans, verification, and review on every change. The gates exist because Charter’s own development depends on them.
Give the work a charter — goal, boundaries, proof — and keep the final say.