MCP tools for agents
Use the compact workflow catalog first, then discover the exact authorized granular action for the job.
Prism is the system of record. An MCP-compatible client should orient, retrieve exact evidence, make reviewable changes only when authorized, and leave irreversible or human-judgment decisions in Prism.
Workflow
Primary steps
- 1Call get_prism_workspace_context, then verify_prism_host_capabilities if the host may have a stale tool snapshot.
- 2Use get_prism_task_context or get_agent_context_pack to understand the task, Project, and evidence before selecting actions.
- 3Use a named workflow tool where one fits; otherwise call discover_prism_actions, inspect the returned schema and scope, then call execute_prism_action.
- 4Preview consequential changes, commit only the approved operation, and report the real outcome and cited evidence back to Prism.
Compact catalog first
A normal owned-Project connection receives 14 workflow tools: get_prism_workspace_context, verify_prism_host_capabilities, discover_prism_actions, execute_prism_action, resolve_prism_object, get_prism_task_context, ingest_research_update, revise_writing_object, manage_project_monitor, generate_project_recap, propose_research_delta, run_research_model, apply_research_bundle, and undo_prism_operation.
The compact catalog is intentional. It gives the agent direct ways to orient, verify host exposure, run common high-level workflows, and reach the full granular registry without guessing tool names or loading a long generic list into every conversation.
Discover and execute a granular action
Call discover_prism_actions only after the agent knows what it needs. Filter for the task and include writes only when the user requested a durable change. Each result includes its authorized scope, JSON input shape, read-only or destructive hint, suggested next tools, and a preview-then-commit protocol when relevant.
Call execute_prism_action with the selected action and its action-specific payload. Reads execute directly. Consequential writes return an operation preview with an operationId and commitToken; submit those unchanged for commit. Use apply_research_bundle only to commit already-previewed operations, and undo_prism_operation only when Prism reports a safe compensation path and the user explicitly confirms it.
A server-authorized action may still be unavailable in a host with a frozen action snapshot. verify_prism_host_capabilities distinguishes actions currently callable by the host from actions Prism authorizes but the host needs to refresh.
Non-negotiable boundaries
MCP does not publish or schedule posts, send email, submit external forms, click confirmation links, execute brokerage trades, connect brokerage accounts, or apply the canonical thesis. Agents may prepare private drafts, Project content, source-backed proposals, and reviewable tracking updates within their scopes.
Use source chunks and their location metadata as evidence. Context packs, observations, models, workspace memory, and a successful tool call can inform reasoning, but they are not proof by themselves. If retrieval is incomplete or evidenceSufficient is false, say so rather than filling the gap from memory.
Granular action catalog
Every action, grouped by the work it does
Agents discover only the authorized actions for their current connection. This reference names the full Prism registry so builders can recognize each workflow.
Orient, discover, and control work
30 toolsMap the authorized workspace, verify what the host can actually call, resolve references, compile task context, and safely apply or undo logged work.
Projects, sources, and inputs
57 toolsCreate and organize Project material, classify sources, inspect retrieval health, and manage supported web, file, RSS, YouTube, and Site Monitor inputs.
Source retrieval and exact evidence
11 toolsSearch authorized knowledge, read bounded sources or stable chunks, and update only the Project material the user explicitly authorized.
Company research and review
18 toolsRead thesis and claim context, identify evidence gaps, generate a Project recap, run safe model scenarios, and create review-only thesis proposals or Research Deltas.
Research Threads and durable tasks
21 toolsMaintain reviewable hypothesis case files, launch bounded Prism research work, and resume durable campaigns or tasks without treating completion as human approval.
Writing Studio
8 toolsRead, create, revise, cite, and add visuals to private writing. Draft and published-work revisions use explicit reviewable flows; MCP never publishes or schedules a post.
Visual artifacts
13 toolsPlan, prepare, validate, save, retrieve, and place structured charts, tables, metric cards, Markdown, HTML, and React artifacts.
Knowledge graph
8 toolsInspect evidence connections and create, revise, archive, delete, or clean explicit Project relationships under the graph scope.
Project Canvas
4 toolsRead, create, and update Project Canvases while preserving authored layout and using the latest version for changes.
Watchlists, monitors, and briefings
18 toolsTrack Market tickers and structured Project conditions, record source-backed updates, and operate reviewable Project monitors and briefings.
Research calendar
5 toolsRead, create, revise, and, with explicit confirmation, delete personal or Project events with dates, sources, and outcome history.
Prism Email
14 toolsRead authorized inboxes, organize messages and labels, manage sender rules, create allowed addresses, and save selected received material into a Project. Email remains inbound-only through MCP.
Workspace memory
11 toolsInspect explainable workspace memory, then explicitly remember, correct, or forget a stable memory record without changing source material.
Portfolio and decision journal
12 toolsRead or update Prism's own tracking records and decision journal. These tools never connect to a brokerage or execute a trade.
Published work
2 toolsList published posts and inspect their evidence health. Revisions become private drafts for a person to review and republish in Prism.