MCP tools

AI workers, work items, their status and the delegate sessions on them. The lowest role that can hold it: viewer.

• list_workers

  List AI workers

  The organization's AI workers with role, status, level and what they are doing now.

  • Reads only

  Inputs status?

• get_worker

  Get an AI worker

  A worker by id or handle: profile, role and open work.

  • Reads only

  Inputs workerId? handle?

• list_work_items

  List work items

  Work items newest first, filtered by status, owner or text. Pass nextCursor back as cursor for the next page.

  • Reads only

  Inputs status? ownerWorkerId? q? limit? cursor?

• get_work_item

  Get a work item

  A work item by id (wi_…) or key (AGT-12): plan, status, artifacts, dependencies, approvals and escalations.

  • Reads only

  Inputs workItem

• list_delegates

  List delegate sessions on a work item

  The bounded agent sessions a worker ran on the item's slices: backend, status, usage and each one's report.

  • Reads only

  Inputs workItem

• get_workspace_overview

  Get the workspace overview

  Read sanitized organization-level work, worker, approval, escalation, provider and recent activity signals. Pass workerId to include launch readiness for that worker.

  • Reads only

  Inputs workerId?

• render_workspace_overview

  Render the workspace overview

  Render a fresh, authorized workspace snapshot in Agent’s responsive read-only workspace view. Pass workerId to include launch readiness for that worker.

  • Reads only

  Inputs workerId?

• get_spend_savings

  Get ways to cut AI spend

  Ways to cut this organization's AI spend, ranked by monthly saving. Each one has the recorded numbers it starts from, how the estimate was computed, where to change it in Agent and a copy-ready prompt. Also lists unusual spend days and what drove them. Estimates at list prices, not bills; it changes nothing.

  • Reads only

  Inputs days? workerId? handle?

Create work items for AI workers, post in work item threads, and revise specs and add notes to them. The lowest role that can hold it: member.

• assign_work

  Assign work to a worker

  Creates a work item owned by a worker (by id or handle). The worker picks it up on its own and may spend budget, change code and open pull requests within its policies; follow it with get_work_item. Use it when the person asks an AI worker to take something on.

  • Changes things
  • Acts outside or cannot be undone

  Inputs title objective workerId? handle? requirements? priority? labels? parentId? dueAt?

• post_message

  Post a message

  Posts to a work item's thread (waking its worker) or to a thread by id. Mention workers with @handle. It is sent as this connection, not as the person.

  • Changes things
  • Acts outside or cannot be undone

  Inputs body workItem? threadId?

• add_spec_note

  Add a note to a spec

  Adds a comment, question or objection on one block of a spec (by blockId, or a ref such as R3) or on the whole spec, or a reply in a note's thread (parentId). It is recorded as this connection, not as the person.

  • Changes things

  Inputs spec blockId? ref? kind body quote? parentId?

• revise_spec

  Revise a spec

  Saves one revision of a spec from block upserts (an id changes that block, only in the fields given; no id adds one) and removals, with a summary. Reply to notes by id in replies; list the notes the revision settles in answers.

  • Changes things

  Inputs spec expectedRevision upsert? remove? title? summary answers? replies?

A work item's brief, its checks, its evidence, the organization's knowledge, and its specs and decision log. The lowest role that can hold it: viewer.

• get_work_item_snapshot

  Get a work item workspace snapshot

  Read one tenant-scoped work item with its checks, gates and evidence records. Use this data tool before render_work_item; it never changes work state.

  • Reads only

  Inputs workItem

• get_work_item_context

  Get a work item's brief

  The item's brief as the platform gives it to an agent working on it: objective, constraints, settled decisions, the rules that apply, the checks, the files in scope and what done means. Read it before working on an item.

  • Reads only

  Inputs workItem

• list_checks

  List a work item's checks

  The item's executable acceptance checks: what each runs, which criteria it covers and how it last ran, plus criteria no check covers yet.

  • Reads only

  Inputs workItem

• list_evidence

  List a work item's evidence

  The item's evidence record: screenshots, recordings, logs, reports and links the platform captured, and waivers.

  • Reads only

  Inputs workItem

• search_knowledge

  Search org knowledge

  Hybrid search over indexed documents (READMEs, ADRs, wiki pages) the workers can see.

  • Reads only

  Inputs text limit?

• get_work_launch_readiness

  Check work launch readiness

  Check the existing execution, provider, queue and budget readiness facts for one worker. This is a read-only preflight and does not reserve spend or launch work.

  • Reads only

  Inputs workerId

• get_work_item_trace

  Trace a work item

  Read a bounded, redacted work item trace with event entries and run counts. Pass nextCursor to continue; this does not mutate runs or approvals.

  • Reads only

  Inputs workItem runId? limit? cursor?

• explain_action

  Explain an action

  Explain one authorized action with its policy events, approval, same-run decisions and context manifest. The response is bounded and redacted.

  • Reads only

  Inputs actionId

• read_evidence

  Read permitted evidence

  Read a permitted content-addressed evidence blob by its exact lowercase SHA-256 hash. Access is checked against the current tenant and caller before any bytes are read.

  • Reads only

  Inputs sha256

• render_work_item

  Render a work item workspace

  Render a fresh, authorized work item snapshot in Agent’s work, check, gate and evidence view. Pass the work item id or key from get_work_item_snapshot.

  • Reads only

  Inputs workItem

• get_repository_atlas

  Get a repository atlas

  A repository's atlas: its stack, commands, standards, infrastructure, data, observability, configuration (variable names only), CI, architecture and findings, computed from its files and Git history without a model. Without repository, lists every profiled repository in one line each. Read it before working in a repository.

  • Reads only

  Inputs repository? section? depth?

• list_specs

  List specs

  Specs, most recently changed first, with status, revision, requirements proven, open notes and the latest audit. Archived specs only when asked for by status. Pass nextCursor back as cursor for the next page.

  • Reads only

  Inputs status? q? limit? cursor?

• get_spec

  Get a spec

  A spec by key (SPEC-7) or id: its blocks with ids and refs, every note, the latest audit, the requirement trace, revisions and the standing decisions it should respect.

  • Reads only

  Inputs spec revision?

• search_record

  Search specs and decisions

  Searches spec titles, blocks, notes and the decision log. A key such as SPEC-7, SPEC-7 R3 or DEC-3 matches exactly first.

  • Reads only

  Inputs q kinds? limit?

The facts, procedures and episodes AI workers remember. The lowest role that can hold it: viewer.

• search_memory

  Search worker memory

  Hybrid recall over the org memory the workers use: facts, procedures, episodes, with why each matched.

  • Reads only

  Inputs text workItem? limit?

• get_memory_record

  Read a memory record

  Read exactly one authorized memory record. Current records are returned by default; set history to inspect superseded, archived or expired history after explicit review.

  • Reads only

  Inputs memoryId history?

Actions waiting for a person. Deciding them stays in Agent. The lowest role that can hold it: viewer.

• list_pending_approvals

  List pending approvals

  Actions waiting for a human decision. Read-only: approving or denying happens in the control plane, by a person.

  • Reads only

  Inputs workItem? limit?

Take work Agent hands to your own Claude or ChatGPT, and report the results. The lowest role that can hold it: member.

• list_my_runs

  List your runs

  Lists your connected-client runs. Overdue open runs are marked expired and their plan-use records are finalized as expired; those runs can no longer report results. Other open runs wait for submit_run_result. This operation updates run state without starting a worker or approving an action.

  • Changes things
  • Acts outside or cannot be undone
  • Needs a person, not a token
  • Safe to repeat

  Inputs status? limit?

• get_run

  Get a run and its brief

  Gets your run and its brief. A fired or uncertain run is marked running with a pickup time; an overdue run is marked expired and can no longer report results, with its plan-use record finalized as expired. Use it to pick up an explicitly selected run, then report with submit_run_result. It does not approve actions or verify completion.

  • Changes things
  • Acts outside or cannot be undone
  • Needs a person, not a token
  • Safe to repeat

  Inputs runId

• start_run

  Start working on a work item

  Start working on a work item yourself, under your own plan: returns a run and the brief. Calling it again for the same item returns the open run. Report with submit_run_result.

  • Changes things
  • Needs a person, not a token
  • Safe to repeat

  Inputs workItem note?

• submit_run_result

  Report a run result

  Report a run once: the outcome, a summary, links to what changed (pull requests, commits, branches) and the checks you ran. It is posted to the item's thread and the owning worker reviews it; the platform runs its own checks before relying on yours.

  • Changes things
  • Acts outside or cannot be undone
  • Needs a person, not a token

  Inputs runId outcome summary links? checks? questions?