Safer retries
Supported write operations accept an idempotency_key, helping agents retry after a timeout without creating the same record twice.
For Agents & Builders
Stop your agents from starting blind.
Connect compatible agents through MCP or the CLI. They can retrieve the project state, evidence, relationships, and constraints behind the work—then return results to the same workspace instead of leaving context trapped in a transcript.
CLI Installation
Start with MCP or the CLI. Both connect compatible agents to the same workspace context.
$ sift projects context execufunction project execufunction context tasks 4 · signals 3 · notes 12 · members 4 people 2 reviewers on this work $ sift work list --status ready WRK-118 agents page: wire live context do_now WRK-121 verify CLI install docs do_next $ sift work claim WRK-118 claimed · project + people + tasks attached verificationCommands: npm test -- --runInBand $ sift work complete WRK-118 --result "agents page context shipped" ✓ verification passed · artifactRefs: 2 → queued for human review
Connected work context
Give agents the state behind the assignment
| Primitive | Read | Write | Examples |
|---|---|---|---|
| tasks | ✓ | ✓ | List by status, project, priority. Create, update, complete. |
| events | ✓ | ✓ | Calendar queries by date range. Create, update events. |
| people | ✓ | ✓ | Search contacts, get org relationships, interaction history. |
| projects | ✓ | ✓ | Full project context: tasks, signals, notes, members. |
| knowledge | ✓ | ✓ | Search notes, get documents. Create notes, upload docs. |
| code | ✓ | — | Semantic code search, git history, blame, expertise, conventions. |
| memory | ✓ | — | User preferences, learned attributes, behavioral patterns. |
Boundaries and evidence
Controls for work that should not run on blind trust.
Permission scoping
Give each personal access token only the read and write scopes it needs. Keep tasks, work items, projects, knowledge, code, people, and other domains independently bounded.
Human review for bounded work
Executable work items can return a summary, artifacts, and verification evidence for review. The agent does the bounded work; a person keeps the judgment and final decision.
Evidence attached to the work
Keep ownership, status, result summaries, artifact references, and verification results with the work item instead of reconstructing the run from a chat transcript.
Cost bounds for managed runs
Managed operator runs can carry explicit USD cost limits and stop when the run reaches its bound. This control applies to managed runs, not every external MCP client.
One workspace across interfaces
MCP and the CLI operate on the same workspace state, so changing interfaces does not create another disconnected copy of the work.
MCP Quickstart
One workspace across interfaces
MCP and the CLI operate on the same workspace state, so changing interfaces does not create another disconnected copy of the work.
MCP Quickstart
Start with MCP or the CLI. Both connect compatible agents to the same workspace context.
Give each personal access token only the read and write scopes it needs. Keep tasks, work items, projects, knowledge, code, people, and other domains independently bounded.
Give your agent a workspace
before you give it instructions.
Start with MCP or the CLI. Both connect compatible agents to the same workspace context.
Executable work items can return a summary, artifacts, and verification evidence for review. The agent does the bounded work; a person keeps the judgment and final decision.