Sequence · v0.1.8.1 · Windows, macOS, Linux
All your favorite harness pieces in one app.
Any model, any agent
Run Claude Code, Codex or a local model (Ollama, or any server that speaks the OpenAI API format) on the same repo.
Draw
A whiteboard, diagrams that render in the Files preview, and an architecture map where each solid line points to its file and line.
Agent workflows
The Agents tab shows context and output for each run, workflow runs and their worktrees. A turn cut off by a restart offers Continue.
Save tokens
A question is answered from one slice of the map, not from every matching file. The measured figures are below.
Small helper models
Small, fast models on your computer recognise what you draw and pick which files to open.
Learns as you work
When the same kind of question fails twice, it proposes a short note for the repo, with the failed runs as evidence; accepted notes go into later prompts. The file finder retrains on the files the agent actually changed.
Pasting code doesn't scale
AI agents learn a codebase by stuffing source into their context. That grows with the repo, misses files buried deep, and diagrams drift from the code anyway.
How it works
Each step below is one clip. Clips use a sample shop repository; the app and scan are real, model answers are sample replies.
01
Read the architecture map
It scans the code and draws the services, databases and topics. Each solid line points to the file and line it came from. One menu switches between Map, Grid and Hotspots.
02
Draw on the whiteboard
Ask for a diagram in chat and the whiteboard draws it laid out, text inside the boxes. Select is the default tool and Space pans.
03
Agents and workflows
Run Claude Code or Codex on the repo. The Agents tab shows context and output for each run, workflow runs and the worktrees they use. Run history and token figures in this clip are illustrative.
Before and after
Each change was tested on a set of questions frozen before anything was scored, and kept only if it won by more than the noise.
Limits
Its earlier file picker lost to grep (a plain text search over the files). The trained model that replaced it finds a file the real fix changed in its top five picks 85% of the time, against 51% for grep, on 8 repositories it had never seen.
Architecture answers, same local agent
Architecture answers, same local agent. Before 0.09, After 0.21. Higher is better.
Answer score from 0 to 1: how closely the files the agent names match hand-checked answers. 36 architecture questions it had never seen, run twice each way.. Source: Sequence report, section 6.
Full answers on the release question bank
Full answers on the release question bank. Before 39% of 100%, After 90% of 100%. Higher is better.
Share of 72 fixed questions answered in full, before and after the changes.. Source: Sequence report, section 2.
Prompt for one question
Prompt for one question. Before 4.9M, After 1,887. Lower is better.
Tokens the model reads for one question: all 428 matching files, versus one slice of the map.. Source: Sequence report, section 4.
Also in 0.1.4
- Restart and continue. A turn cut off by quitting shows Continue, and queued prompts come back Held.
- Several agents at once. Each gets its own chat and worktree, then you compare and merge one.
- Open a pull request from a worktree chat. Files that look like secrets are held back and listed.
- Bring old chats in. Claude Code and Codex sessions import as chats; the CLI folders are not changed.