Build the whole cycle.
From your terminal.

The AI coding agent for product engineers — discover, build, verify and ship, on any model, in any terminal. Local-first, safe by default, no account.

Apache-2.0
excalibur — m-shell
Add OAuth login to the API
L3 — Implement in branch · kimi-k2.7-code
Plan5 steps
Implementauth/oauth.ts
edit auth/oauth.ts+62 −8
▸ pnpm test authrunning…
Verify
Document
Review
Pull Request
████░░░░1m12s·$0.42·standard-safe·no push

The m-shell

Not commands. A conversation.

Run excalibur and just say what you want. It routes intent to the right workflow, shows you the plan, and works in the open — with a live TUI you can read.

  • Plain language or /commands — it routes intent to the right workflow.
  • Inline approvals: y · N · always, respecting standard-safe.
  • ↑/↓ history, @file and #symbol references, resumable sessions.
  • Watch it think, edit and test live — one event stream, one source of truth.
excalibur — m-shell
add pagination to the orders endpoint
intent: feature · workflow: standard-feature · L3
plan ready — 5 steps
edit api/orders.ts +41 −6
run migration? y · N · always
18 tests pass · docs updated
opened PR #214 — ready to review
Autonomy

Dial it from a question to a full agentic run

Express intent — Excalibur picks the workflow and sizes the work. You decide how far it goes, per task.

less autonomymore autonomy
L0Reviewreview
L1Assistask
L2Propose patchpatch
L3Implement in branchrun
L4Full agenticrun --careful
Local-first

The whole cycle — on your machine

Discovery, a local backlog, runs, replay and a web dashboard — your fleet of agents, no cloud, no account.

Local web dashboard

A lazy daemon serves a fleet view of every run — diffs, tests, cost — at localhost.

Work items & kanban

A local backlog and board — no Linear, no Jira, no cloud required.

Time machine

Rewind any run and fork from any point — only what changed re-runs.

Self-sizing swarm

Big tasks fan out to parallel agents in isolated worktrees — sized for you.

Memory that compounds

Decisions accumulate in .excalibur/memory — every run gets sharper.

Discovery gate

Clarify scope first — and decide whether to build it at all.

Fits your setup

Keep what you already have

Excalibur detects and respects your existing instructions — no rewrites. It asks before enabling anything executable.

  • CLAUDE.mdrespected
  • AGENTS.mdrespected
  • Cursor rulesrespected
  • Copilot instructionsrespected
  • README · docs · ADRsrespected
  • SKILL.mdreview first

Extend everything

YAML, the SDK, MCP & LSP

Shape how your team works with simple YAML — or go deeper with the TypeScript SDK. External tools and real compiler errors flow straight to the agent.

YAML / Markdown

Workflows, methodologies, prompts, safety presets, model routing.

TypeScript SDK

On npm — work-item providers, channels, model adapters, tools, policies.

Custom agents

Self-contained Markdown personas — role, model, tools and permissions per agent.

MCP servers

External tools enter the agent with the manifest’s permissions.

LSP diagnostics

Real compiler errors anchor fixes — no hallucinated lines.

IDE extension

VS Code, Cursor and Windsurf bridge to the agent over ACP.

Runs anywhere you do — any terminal or IDE (JetBrains, Cursor, VS Code, SSH, CI), on macOS, Linux and Windows.

portable by design

Safe by default

Delegate big work without fear

Nothing is modified, applied or pushed without your explicit approval. standard-safe is on from the first command.

Approval gates

Every write, command and push pauses for an explicit yes.

Sandboxed execution

Agents run isolated — no network access by default.

Secrets never leak

.env files and keys are blocked, never read or sent.

Isolated branches

Work lands in dedicated branches — never your tree.