Fleet control
Run a dozen agents at once. Each gets its own worktree and sandbox, so nothing collides. Pause, fork, promote or kill any run without disturbing the others.
Private beta — opening in waves through 2026
Stop babysitting one agent.
Stop babysitting one agent.
Command a fleet.
Commanda is a development environment built for delegation. Brief many agents at once, watch every action as it happens, and take their work as reviewable diffs — not as chat transcripts you have to archaeologically excavate.
Local-first · model-agnostic · your code never leaves your machine by default.
Questions? hi@commanda.ai
The workspace — fleet, plan, diff and timeline in one view. Interface in development.
01 / the premise
A chat window was never an IDE. It was the first thing that fit.
Agents got good at writing code faster than our tools got good at supervising them. So we scroll transcripts, lose track of what changed, run one task at a time, and re-explain the same context every morning.
-
Serial by force
One agent, one branch, one thread of attention — while your machine sits at 6% CPU.
-
Unreviewable output
Diffs buried in prose. You either read everything or trust everything.
-
No way back
When a run goes sideways at step nine, your only rewind is a fresh prompt.
02 / capabilities
Built for the part that's actually hard: staying in control.
Plans you can edit
Every task begins as a plan you can rewrite. Delete a step, add a constraint, pin a file. Agents check in at the boundaries you set instead of improvising past them.
Diff-first review
Work arrives as a changeset with tests attached, not a wall of chat. Comment on a line and the agent revises exactly that — no re-prompting from scratch.
Rewind anything
Every tool call is a checkpoint. Scrub the timeline, branch from step three, and run two approaches side by side to see which one survives your test suite.
Context that stays yours
Point Commanda at repos, docs, tickets and runbooks. Indexing happens locally, and each agent gets a scoped slice — not your entire filesystem.
Any model, any endpoint
Frontier APIs, open weights, or your company's gateway. Choose per agent, per step — and switch mid-project without rewriting a thing.
12concurrent agent runs on a developer laptop, each fully isolated.
< 1sto fork any run into a parallel branch from any checkpoint.
0required cloud dependencies. Bring your own keys, or your own weights.
Design targets for the 1.0 release, measured on a 2024-class 8-core laptop.
03 / the loop
Brief. Fan out. Review. Merge.
The whole environment is shaped around one loop, and every part of it is interruptible.
STEP 01
Brief in plain language
Describe the outcome, drop in a ticket, or select a failing test. Commanda drafts a plan with the files it intends to touch and the checks it intends to run. You approve it, edit it, or throw it out.
STEP 02
Fan out into isolated runs
Split one brief across several agents, or launch several briefs at once. Each run lands in its own git worktree inside a sandbox with declared network and filesystem scope. Nothing touches your working tree until you say so.
STEP 03
Supervise at a glance, intervene in one keystroke
The fleet view shows what every agent is doing right now — the file, the command, the test result. Step in to correct a run mid-flight, or let it finish and read the diff. Ambient status, never a notification storm.
STEP 04
Review as code, merge with confidence
Accept hunk by hunk, request a revision on a single line, or fork the run and compare two solutions. What lands is a clean commit with the plan, the trace and the test output attached to it.
04 / the feel of it
An interface that gets out of your way.
Supervising agents is a new kind of work, and it deserves better than a sidebar bolted onto a text editor. Commanda is designed for the rhythm of delegation: long stretches of glancing, short bursts of precision.
- Keyboard first
One palette reaches every agent, run, diff and checkpoint. Nothing important hides behind a mouse.
- Calm by default
Status is ambient — colour, motion, position. No badge storms, no modal interruptions, no sound.
- Legible density
Twelve runs on one screen without squinting. Type, spacing and colour do the work that borders usually do.
- Your editor stays yours
Open any run in VS Code, JetBrains or Neovim at the exact file and line. Commanda directs the work; it doesn't demand your muscle memory.
# the same environment, without the windows commanda run "add sliding-window rate limits" \ --blueprint backend \ --parallel 3 \ --plan-first # review what came back commanda diff --agent limiter --stat src/middleware/limiter.ts +38 src/router/public.ts +6 -2 test/limiter.spec.ts +74 # ship the one you liked commanda merge limiter --squash
Everything in the UI is a command, and every command is scriptable — including in CI.
name: reviewer model: $YOUR_GATEWAY/any-model # or a local endpoint context: include: ["src/**", "docs/adr/**"] exclude: ["**/*.env", "infra/secrets/**"] tools: shell: { allow: ["pnpm test", "pnpm lint"] } network: deny mcp: ["linear", "sentry"] hooks: before_edit: ./scripts/guard-migrations.sh after_run: ./scripts/post-to-slack.sh policy: require_plan_approval: true max_files_changed: 40
05 / customization & integration
Your workflow, encoded — not ours, imposed.
An agent that doesn't know your conventions is a liability. Commanda makes those rules explicit, versioned in your repo, and reviewable like any other code.
- Blueprints
Declare an agent in one YAML file: model, context scope, permitted tools, guardrails. Commit it, review it, share it with the team.
- Hooks
Run your own scripts before or after any tool call. Block a migration, enforce a codeowner, post to Slack, fail the run.
- Skills
A folder of markdown and scripts your agents load on demand — house style, deploy runbooks, the tribal knowledge that lives in someone's head.
- MCP & APIs
Connect any Model Context Protocol server in a line. Tickets, logs, dashboards and databases become first-class context.
- CI-native
The same blueprints run headless in your pipeline, so the agent that reviews your PR is the one you tuned on your laptop.
Nothing runs unscoped
Every agent executes inside a container with declared filesystem and network permissions. Deny-by-default is the default.
Your code stays put
Indexing, history and traces live on your machine. The only thing that leaves is what the model you chose needs to see.
Every action, on the record
Full traces of prompts, tools, commands and diffs — exportable, greppable, and attached to the commit that shipped.
06 / early access
Get in while it's still being shaped.
Commanda is in active development. We're onboarding a small number of teams at a time and building against their real repositories — messy monoliths especially welcome.
- What you get
Early builds, a direct line to the people writing the code, and free use throughout the beta.
- What we ask
Honest reactions. Tell us where it slows you down and we'll fix it before we charge anyone.
- Prefer email?
Write to hi@commanda.ai — a person reads every message.
hi@commanda.ai