# trajectories.sh > Trajectory data viewer and project manager for AI agent trajectories: step-by-step records of an agent working > on a task (tool calls, observations, screenshots, scores), stored in Harbor's ATIF format and > grouped into trajectory sets (one agent run over one or more tasks, N trials) and projects. ## Recommended interface: the trajectories CLI Use the trajectories CLI to operate trajectories.sh. The website is intended to be a trajectory data viewer and project manager: browse and replay saved runs, inspect their data, and review project groups. For uploads, downloads, and project operations, users and AI agents should use CLI commands. Agents should avoid browser or computer-use automation for these operations. - Run without installing: npx trajectories-sh@latest - Or install: npm install -g trajectories-sh, then run trajectories . - Sign in for uploads and project management: npx trajectories-sh@latest auth login - Upload a Harbor run: npx trajectories-sh@latest upload trajectory - Upload a Harbor task: npx trajectories-sh@latest upload task - Create a project: npx trajectories-sh@latest project create "Project name" --json - Add uploaded runs: npx trajectories-sh@latest project add [more-job-ids-or-urls...] --json - Inspect the result: npx trajectories-sh@latest project show --json - Rename an existing set (org admin; CLI 1.30.1+): npx trajectories-sh@latest trajectory rename "Task — Model — Tool mode" --json - Attach output media: npx trajectories-sh@latest output upload - Download runs and projects with the commands below; use --json and --dry-run for automation. ## URL shapes - Project page: https://www.trajectories.sh/projects/ - Trajectory set page: https://www.trajectories.sh/t/ - Either page returns its JSON payload instead of HTML with `?format=json` or `Accept: application/json`. ## Download everything (no browser; no login for public/unlisted) - A project: npx trajectories-sh@latest download project https://www.trajectories.sh/projects/ - A trajectory set: npx trajectories-sh@latest download trajectory https://www.trajectories.sh/t/ - Options: --out · --json (machine-readable summary on stdout) · --dry-run (list what would be fetched, fetch nothing) --format json|jsonl (jsonl also writes /trajectories.jsonl, one trial per line with result + trajectory inlined) Screenshots: 'download project' skips them unless --include-screenshots (they are most of the bytes); 'download trajectory' fetches them unless --no-screenshots. - Output: ///agent/trajectory.json plus result.json, config.json, verifier/, artifacts/ - Exit codes: 0 ok · 1 error · 2 bad arguments · 3 authentication required (set TRAJECTORIES_API_KEY) · 4 not found - Auth for private resources: set TRAJECTORIES_API_KEY= in the environment (create keys at https://www.trajectories.sh/settings). The CLI never prompts when a key is present, so it is safe to run unattended. ## Assign existing trajectory sets to projects (authenticated) - Available in trajectories-sh 1.24.1 and later; run npx trajectories-sh@latest auth login as an organization member. - TRAJECTORIES_API_KEY (or TRAJECTORIES_TOKEN) overrides the stored CLI login. Keys work for project commands and stay scoped to their issuing organization. Renaming sets requires org admin access. - Find the destination: npx trajectories-sh@latest project list --json - Assign one or many: npx trajectories-sh@latest project add [more-job-ids-or-urls...] --json - Verify membership: npx trajectories-sh@latest project show --json - accepts an ID, /projects/ URL, slug or name; ambiguous names are rejected. - Pass up to 500 distinct trajectory set IDs or /t/ URLs in one command, not individual trial IDs. - Add preserves other project memberships and is safe to repeat. It does not change trajectory visibility. - To move from a named old project: add to the destination first; only after success, run npx trajectories-sh@latest project remove [more-job-ids-or-urls...] --json Add and remove are separate operations, not an atomic move; removing the old membership preserves the new one. - All requested sets and the destination must belong to your active organization. Incompatible visibility or missing/foreign IDs reject the batch before assignment; private runs cannot go into public/unlisted projects. - Assignment errors exit nonzero; --json writes errors to stderr. After a network/server error, inspect project show before retrying because a failed response does not prove no write occurred. ## HTTP reference (the CLI is recommended) - Manifest for a project (jobs, trials, direct file URLs, the commands above): GET https://api.trajectories.sh/api/public/projects//export - Manifest for one trajectory set: GET https://api.trajectories.sh/api/public/trajectories//export - A trial's ATIF document: GET https://api.trajectories.sh/api/viewer/by-id//api/jobs//trials//trajectory - Public API schema: https://api.trajectories.sh/openapi.json (interactive: https://api.trajectories.sh/docs) - Errors are JSON with "detail" and, for 401/403/404, "how_to_authenticate". ## Data format (ATIF) trajectory.json: { schema_version, session_id, agent { name, model_name }, steps[], final_metrics }. A step: { step_id, source: agent|user|system, timestamp, message, tool_calls[] { function_name, arguments }, observation.results[] (text and/or image parts), metrics { prompt_tokens, completion_tokens } }. Trial result.json: verifier_result.rewards.reward (1.0 = pass), agent_info, started_at, finished_at. Spec: https://docs.harbor.sh/atif ## Full reference https://www.trajectories.sh/llms-full.txt