One harness. Any model. Every agent.
Byte Poet is not a chat box with a code theme. It is a harness: a loop that plans, acts, reads what came back, and corrects — wrapped in the gates, checkpoints and permissions that make letting a model act on real files reasonable.
The loop, in order
The same four beats run under every agent. What changes between agents is the toolset, not the discipline.
-
Reason
The agent plans against a budgeted context window — retrieved code, conversation history, project instructions and its own notes, assembled to fit the model it is actually running on.
-
Act
It calls a real tool: edit a file, run a command, search the repo, open a page, query a language server. Read-only tools can fan out in parallel; anything that writes is serialised.
-
Observe
Output comes back as ground truth — exit codes, diagnostics, test results, diffs, screenshots — not as the model's guess about what probably happened.
-
Correct
It reconciles plan against reality and goes again, until the work compiles, the tests run, and the task is actually done.
What the agents can actually do
Every capability below is a tool the model can call, a gate it must pass, or a surface you can watch it work in.
Workspace & code
- Read, write, exact-match replace, atomic multi-edit and per-file undo
- Full-text and semantic search over the workspace
- Language-server navigation and post-edit diagnostics
- Git as a first-class tool — branches, diffs, commits
- Isolated git worktrees for risky refactors, merged back on success
Running things
- A real terminal with streaming output and background jobs
- Live app preview — the agent starts your app and drives it
- A browser it can navigate, read and screenshot
- Build and test verification gates before work is called done
- Auto-format on save and regression checks on every edit
Judgement & control
- Plan mode — read-only exploration before anything is written
- A command risk classifier that gates what auto-approves
- Per-workspace permission rules you set once and reuse
- Hooks that can observe, or block, an action before it runs
- Checkpoints on every step, with one-click rewind
Scale & memory
- Sub-agent delegation for work that deserves its own context
- A task queue the agent plans against and works down
- Twenty built-in skills covering discovery through deploy handoff
- Retrieval, sliding-window memory and token-aware compaction
- MCP servers, so your own tools join the agent's toolbox
Made to be watched
An agent you cannot see is an agent you cannot trust. Every run is legible while it happens.
Reasoning, tool calls, arguments, exit codes and durations, token by token, as they happen.
Explorer, editor and agent side by side — with diffs, problems, terminal and preview in the same window.
Type a correction while it works and it folds into the next turn. Stop it and the partial answer is still saved.
Tokens and cost recorded on every exit path — including runs that error out or hit a limit.
Plugs into what you already run
One OpenAI-compatible route to Claude, Gemini and OpenAI. The proxy stores no credentials — your key rides with the request.
One key, hundreds of models, called directly with your credentials. The catalogue and vision-capability flags are read live from OpenRouter.
Any OpenAI-compatible local endpoint. Small models get a conservative capability profile automatically; large ones get the full tool loop.
Point a webhook at a workspace and an issue can start an agent run, on a branch, with a pull request at the end.
Configure Model Context Protocol servers per workspace and their tools appear alongside the built-ins.
A full VS Code in the browser sits beside the agent, on the same workspace files.
See it run on your own repository.
Enter the studioByte Poet is in private beta. Sign-in is limited to allowlisted accounts. Request access