Upload from any agent
Publish traces from 10+ supported agents directly in the CLI.
Share and collaborate on your coding agent sessions. Made for teams, and free to get started.
Download the CLI to get started
Works with your favourite agents






Start sharing traces from any agent, in minutes.
Publish traces from 10+ supported agents directly in the CLI.
Copy the link & share your trace with your team or the world.
UserMenu because it uses imperative WAAPI and can re-measure on state changes. I'm reading that implementation next and then I'll verify smoothing techniques from current guidance.Download a full trace or continue working on someone else's.
Download a trace, or continue it in Claude Code. The open agent menu lists OpenCode, Pi, Claude Code, Codex, Cursor, Amp, Cline, OpenClaw, GitHub Copilot, Hermes, Droid, Antigravity, with Claude Code selected.
Traces for Teams
Traces gives teams one place to see work in progress, share context on finished work, and see which agents people are using.
Create your Team

“Github increasingly doesn't feel like the best place to understand the work done on a codebase. Agent traces provide a much more human-readable overview. Just started using traces.com. Feels quite nice.”
Millin Gabani, CEO of Worklayer
See original postShare agent conversations without worrying about sensitive data.
Share your traces privately, directly, or publicly, so only the right people see them.
maujim · 102 messages
Diagnose and Enhance Voltage Error Reporting
maujim · 102 messages
Diagnose and Enhance Voltage Error Reporting
maujim · 102 messages
Diagnose and Enhance Voltage Error Reporting
Set team-level policies to control how your team can share traces.
Traces can be published as:
We automatically strip sensitive data like API keys, emails & database credentials from traces on publish.
Shared the rollout trace after scrubbing [REDACTED], [REDACTED], and [REDACTED] from the assistant reply before sending the link to the team.
The published run keeps the reasoning intact while replacing keys, customer emails, and database URLs with clear [REDACTED] markers anyone can spot immediately.
Reviewers still understand what happened, but the sensitive values stay hidden behind [REDACTED] in every shared view.
Start simple & integrate more as you go.
Publish & manage traces directly from your terminal.
Let your agent share traces as you work.
Publish traces via API with your own tools, schedulers & workflows.
Automatically share traces on every commit from your CI/CD pipelines.
Free for individuals & small teams, flexible at scale.
Core
Free
Custom
Get in touch
See what ships, loop in the right agents, and share traces with git hooks & custom skills.
Team analytics preview showing Claude Code, Codex, Cursor, Droid, and Amp as the top agents, with an average session length of 47 minutes and 82.0 percent AI output.
A GitHub-style pull request timeline shows Maya Chen committing a docs update, a Traces bot comment with pull request trace links, two preview deployments, and a pull request mention.
Traces found for this PR:
Team member list showing Maya Chen as admin, Theo Brooks as member, Alice as the invited agent, Lina Park as member, and Ari Singh as member.

A terminal conversation showing a request to share a trace, the Traces skill that ran, and the private share link confirmation.
Great, I added back the function.
Share this trace
Skill Traces
Ran traces share --trace-id w9g0svsjx839ngt --json
Shared your trace to https://traces.com/s/w9g0svsjx839ngt as private.
Supported agents include Claude Code, Cursor, OpenCode, Codex, Pi, Amp, Cline, OpenClaw, GitHub Copilot, Hermes, Droid, Antigravity.
Install the CLI to get started or make an account first.
Browse traces & learn how people are using agents. Contribute by sharing your own traces.
Browse public tracesThe user requested completing a detailed multi-step scenario for adapter-quality dataset capture using specific tools and constraints. The assistant executed the steps in order, including file globbing, creating and modifying scratch files, and verifying intermediate outputs while adhering to mutation restrictions.
The session involved inspecting a missing file path to capture the exact error without modifying files, followed by reading and summarizing the adapter-quality flow from a README. The assistant requested approval before creating a harmless temp file, which was denied, and then triggered a read-only subagent to test parsing of various metadata. Finally, the assistant confirmed that no files were modified during the session and nothing needed to be reverted.
The user requested a simple 2D runner game called Super Dario, inspired by Super Mario, implemented entirely in raw JavaScript, CSS, and HTML within a single file. The assistant built the game with creative references to Anthropic and technical staff, initialized a git repository, pushed the code to the user's GitHub repo, enabled GitHub Pages, and confirmed the game is live and playable.
The user conducted a multi-turn session focused on adapter-quality capture tasks within a CLI project. The assistant listed adapter files, created and edited a scratch markdown file, ran shell commands including an expected failure, performed lint checks, managed todo items, conducted a web search on Convex scheduled functions, and finally deleted the scratch file while listing normalized TraceEvent types from memory.
The user continued a multi-turn session focused on adapter-quality capture tasks. The assistant listed adapter file paths, created and edited a markdown file, executed shell commands reporting outputs including an expected error, ran a linter with no issues found, noted the absence of a todo tool and web search capabilities, and finally deleted the temporary file while providing a reasoned answer about TraceEvent types.
The user reported a UI issue where a connection appeared "working" despite the OAuth handshake never completing. The assistant investigated the backend and UI logic, identified a false-positive in the UI status interpretation, implemented a comprehensive fix including UI and backend reconciliation, added tests, and pushed the changes to a branch. The fix is already included in an open PR (#23) with additional follow-up commits.
The user reported that the Today page reloads data from scratch on every visit, causing a poor user experience. The assistant investigated and identified two main issues: the client refetches data on every mount without caching, and the server runs an expensive retrieval pipeline on every request without caching. The assistant implemented a two-part fix involving a server-side Redis cache and a client-side stale-while-revalidate cache keyed by user ID to prevent cross-tenant data leaks.
The session addressed fixing and improving the relationship graph to eliminate false relationships, duplicates, and junk entities. The assistant investigated root causes, implemented extraction and filtering fixes, enhanced the strength model, updated the UI to present a you-centered view, and ensured the feature was verified with live data. Later, the assistant identified that the branch containing the relationship-intelligence work was merged but the commits were pushed afterward without an open PR, leading to a cleanup and rebasing effort to create a clean, mergeable PR.
The user requested completing a detailed multi-step scenario for adapter-quality dataset capture using specific tools and constraints. The assistant executed the steps in order, including file globbing, creating and modifying scratch files, and verifying intermediate outputs while adhering to mutation restrictions.
The session involved inspecting a missing file path to capture the exact error without modifying files, followed by reading and summarizing the adapter-quality flow from a README. The assistant requested approval before creating a harmless temp file, which was denied, and then triggered a read-only subagent to test parsing of various metadata. Finally, the assistant confirmed that no files were modified during the session and nothing needed to be reverted.
The user requested a simple 2D runner game called Super Dario, inspired by Super Mario, implemented entirely in raw JavaScript, CSS, and HTML within a single file. The assistant built the game with creative references to Anthropic and technical staff, initialized a git repository, pushed the code to the user's GitHub repo, enabled GitHub Pages, and confirmed the game is live and playable.
The user conducted a multi-turn session focused on adapter-quality capture tasks within a CLI project. The assistant listed adapter files, created and edited a scratch markdown file, ran shell commands including an expected failure, performed lint checks, managed todo items, conducted a web search on Convex scheduled functions, and finally deleted the scratch file while listing normalized TraceEvent types from memory.
The user continued a multi-turn session focused on adapter-quality capture tasks. The assistant listed adapter file paths, created and edited a markdown file, executed shell commands reporting outputs including an expected error, ran a linter with no issues found, noted the absence of a todo tool and web search capabilities, and finally deleted the temporary file while providing a reasoned answer about TraceEvent types.
The user reported a UI issue where a connection appeared "working" despite the OAuth handshake never completing. The assistant investigated the backend and UI logic, identified a false-positive in the UI status interpretation, implemented a comprehensive fix including UI and backend reconciliation, added tests, and pushed the changes to a branch. The fix is already included in an open PR (#23) with additional follow-up commits.
The user reported that the Today page reloads data from scratch on every visit, causing a poor user experience. The assistant investigated and identified two main issues: the client refetches data on every mount without caching, and the server runs an expensive retrieval pipeline on every request without caching. The assistant implemented a two-part fix involving a server-side Redis cache and a client-side stale-while-revalidate cache keyed by user ID to prevent cross-tenant data leaks.
The session addressed fixing and improving the relationship graph to eliminate false relationships, duplicates, and junk entities. The assistant investigated root causes, implemented extraction and filtering fixes, enhanced the strength model, updated the UI to present a you-centered view, and ensured the feature was verified with live data. Later, the assistant identified that the branch containing the relationship-intelligence work was merged but the commits were pushed afterward without an open PR, leading to a cleanup and rebasing effort to create a clean, mergeable PR.