Project memory
Pro and upA vault at .firstbeam/memory, committed with the repo. Notes are typed — fact, decision, pattern, context, lesson, warning, todo, reference — and linked wiki-style. One switch injects them into chats.
Live interface · simulationOpen full screen
Coordinatororchestrator
Goal split into two tracks — auth touches three modules.
Scoutdiscovery
Repo scanned. Session store and the cookie helper are the hidden deps.
Builderimplementation
Track A is mine: token refresh. Tests first.
Reviewerquality
Blocking track B — it renames the cookie. Keep the old name one release.
Builderimplementation
Good catch. Aliasing it, pushing again.
Coordinatororchestrator
01 Agent mode
Agents you name, not sessions you juggle. Lead, analyst, developer, QA and designer share one thread: the work passes from hand to hand, and every step — who did what, who they handed it to, which files they touched — lands in the same transcript.
Give each agent a name, a face, an engine and a permission level. The team is yours and it persists across threads.
Write the goal once. Lead routes it — you are not picking the next agent at every step.
Every turn ends in a handoff or a done. With auto-handoff on, the next turn starts by itself, up to the hop cap you set.
Permissions are real: an agent declared read-only cannot write. Cost lands per turn, and the budget stops an unattended chain.
02 Chat and terminal
Both run inside the project you already opened — same repository, same engines, same memory. Chat holds the thread and the context; the terminal gives you a real shell. Moving between them costs you no re-explaining.
Add a pricing page — reuse the existing Pricing component.
Touching three files. Route is up, component next.
Hold on — make the plan fit the enterprise tier too.
Got it — steering the running turn now.
steered mid-turnA message sent mid-turn steers the running turn instead of waiting in line behind it.
Terminal panes are actual PTYs: scroll back, resume, run whatever the shell runs.
Chat embeds the project brief in every turn; a terminal pane syncs the vault and points the engine at it.
03 Around the agent
The modes are in 01–02. This is the layer that makes a run cheaper, stops it at the right moment, and keeps what it learned.
A vault at .firstbeam/memory, committed with the repo. Notes are typed — fact, decision, pattern, context, lesson, warning, todo, reference — and linked wiki-style. One switch injects them into chats.
Portable instruction packs. Drop one onto a chat, a terminal pane or a lab card and it installs into the project, so every engine picks it up.
One kanban.json in the repo. A card opens a chat with itself already in the composer, or is pushed into a terminal pane.
Notes and freehand sketches beside the code in notes.json. Send to chat and add to terminal move the page into work.
Read-only, supervised or auto, plus a whitelist that stays auto-approved while supervised. Plan turns are never widened by the switch.
A message sent during a turn steers that turn, or joins the queue. When steering is not possible it falls back to the queue on its own.
A browser docked beside the conversation. You watch the same panel the assistant drives. Read-only still refuses clicking and typing.
Goals, tasks, reviews and human gates in one graph, each step with its own engine, model and effort.
A second switch lets agents write notes themselves. The panel searches the vault, graphs how notes link, and reports an estimated token saving from reuse.
Which chats are working, which are blocked on you, which terminals are busy. It stores what happened and how it turned out, never the conversation. It never leaves the machine.
The catalog states what a plugin will be allowed to do — run commands, read project files, write into the project — before it installs.
Board, notes, memory and settings are written into .firstbeam/ and travel with ordinary commits. Chat sessions stay at the app level, and do not.
Each engine is its own command-line tool, signed in to its own account. firstbeam never handles a password or a token. A hidden engine is not polled at all.
Start here
You build the product. firstbeam keeps the environment ready.
Download firstbeam — or open the interactive demo first.
Install
Download the desktop build and open it.
Connect the engines you already pay for
Each signs in to its own account. firstbeam never handles a password or a token.
Open a folder
It becomes a project. Board, notes, memory and settings live in .firstbeam/ and travel with your commits.