2921 Smr Sdk MCP Control Proposal 2026 02 19 D3b17850

SMR SDK + MCP Control Proposal (2026-02-19)

Goal

Add first-class Synth SDK and MCP support for controlling Synth Managed Research (SMR) projects and runs.

This proposal covers:

• a new synth_ai.sdk.managed_research client,
• a new MCP server surface that wraps the SDK,
• and rollout sequencing to align backend/frontend/spec route contracts.

Current State (from code)

What is already implemented

• Backend SMR control-plane routes exist in /Users/joshpurtell/Documents/Github/backend/app/api/v1/routes_smr.py.
• SMR data-plane MCP exists in /Users/joshpurtell/Documents/Github/synth-managed-research/crates/smr-worker-host/src/main.rs and serves:
  • POST /mcp/orchestrator
  • POST /mcp/worker/:worker_id
• Orchestrator and worker MCP tool sets are defined in:
  • /Users/joshpurtell/Documents/Github/synth-managed-research/crates/smr-worker-host/src/smr_orchestrator_tools.rs
  • /Users/joshpurtell/Documents/Github/synth-managed-research/crates/smr-worker-host/src/smr_worker_tools.rs

Gaps relevant to SDK + MCP product surface

• synth-ai currently has no dedicated SMR control client module under synth_ai/sdk/.
• Frontend and specs use project-scoped run routes (/smr/projects/{project_id}/runs/...) while backend currently exposes canonical run routes mostly under /smr/runs/....
• Frontend SMR provider-key route currently prefers encrypted payloads (encrypted_key_b64) while backend SMR provider-key schema expects plaintext api_key.

Proposed SDK Surface

Module layout

Add:

• /Users/joshpurtell/Documents/Github/synth-ai/synth_ai/sdk/managed_research/__init__.py
• /Users/joshpurtell/Documents/Github/synth-ai/synth_ai/sdk/managed_research/client.py
• /Users/joshpurtell/Documents/Github/synth-ai/synth_ai/sdk/managed_research/models.py
• /Users/joshpurtell/Documents/Github/synth-ai/synth_ai/sdk/managed_research/errors.py

Export from:

• /Users/joshpurtell/Documents/Github/synth-ai/synth_ai/sdk/__init__.py

Primary class

ManagedResearchClient

Constructor pattern should match existing SDK clients:

• api_key: Optional[str] = None (fallback SYNTH_API_KEY)
• backend_base: Optional[str] = None (fallback Synth backend resolver)
• timeout_seconds: float = 30.0

Method groups

Project lifecycle:

• create_project(...)
• list_projects(include_archived=False)
• get_project(project_id)
• patch_project(project_id, ...)
• get_project_status(project_id)
• pause_project(project_id)
• resume_project(project_id)
• archive_project(project_id)
• unarchive_project(project_id)

Onboarding + keys:

• onboarding_start(project_id)
• onboarding_complete_step(project_id, step, status, detail=None)
• onboarding_dry_run(project_id)
• onboarding_status(project_id)
• set_provider_key(project_id, provider, funding_source, api_key=None, encrypted_key_b64=None, encrypt_before_send=False)
• provider_key_status(project_id, provider, funding_source)

Run controls:

• trigger_run(project_id, timebox_seconds=None)
• list_runs(project_id)
• list_active_runs(project_id)
• get_run(run_id, project_id=None)
• pause_run(run_id)
• resume_run(run_id)
• stop_run(run_id)

Human-in-loop:

• list_project_questions(project_id, status_filter="pending")
• list_run_questions(run_id, project_id=None)
• respond_question(run_id, question_id, response_text, project_id=None)
• list_project_approvals(project_id, status_filter="pending")
• list_run_approvals(run_id, project_id=None)
• approve(run_id, approval_id, comment=None, project_id=None)
• deny(run_id, approval_id, comment=None, project_id=None)

Artifacts + observability:

• list_run_artifacts(run_id, project_id=None)
• get_artifact(artifact_id)
• get_artifact_content_response(artifact_id, disposition="inline", follow_redirects=False)
• get_usage(project_id)
• get_run_spend_entries(run_id) (admin spend-ledger rows)
• get_run_economics(run_id) (admin run economics summary)
• get_run_usage_by_actor(run_id, project_id=None, include_done_tasks=True)
• get_ops_status(project_id, include_done_tasks=None)
• search_victoria_logs(project_id, ...)

Compatibility rule

For all run-scoped read/write methods, use project-scoped route first and fallback to canonical /smr/runs/... if project-scoped returns 404.

This gives a stable SDK surface now while backend aliases are added.

Error model

Define ManagedResearchApiError including:

• status_code
• method
• path
• detail_snippet

This should mirror existing SDK error ergonomics (clear status + endpoint context).

Proposed MCP Surface (in synth-ai)

Why separate MCP

SMR already has internal orchestrator/worker MCP for data-plane runtime. This proposal adds a customer/operator-facing MCP server that exposes control-plane actions via the new SDK client.

Transport

• Default: stdio
• Optional: Streamable HTTP

Rationale: aligns with MCP architecture and transport guidance, where stdio is common local transport and Streamable HTTP is recommended for production deployments.

Tool namespace

Use explicit names to avoid collision with internal worker tools.

Project tools:

• smr.project.create
• smr.project.list
• smr.project.get
• smr.project.update
• smr.project.status
• smr.project.pause
• smr.project.resume
• smr.project.archive

Onboarding tools:

• smr.onboarding.start
• smr.onboarding.complete_step
• smr.onboarding.dry_run
• smr.onboarding.status
• smr.provider_key.set
• smr.provider_key.status

Run tools:

• smr.run.trigger
• smr.run.list
• smr.run.list_active
• smr.run.get
• smr.run.pause
• smr.run.resume
• smr.run.stop

Approval/question tools:

• smr.question.list_project
• smr.question.list_run
• smr.question.respond
• smr.approval.list_project
• smr.approval.list_run
• smr.approval.approve
• smr.approval.deny

Artifact/ops tools:

• smr.artifact.list_run
• smr.artifact.get
• smr.artifact.content_link
• smr.usage.get
• smr.run.spend_entries.get (admin scope)
• smr.run.usage_by_actor.get
• smr.ops_status.get
• smr.logs.search

Safety defaults

• No /smr/internal/* tools in this MCP.
• Explicitly mark side-effecting tools in tool descriptions.
• Require required IDs (project_id, run_id, etc.) and reject ambiguous calls.
• Return structured JSON only (no markdown blobs) for machine composability.

Packaging + CLI integration

Add optional dependency group in pyproject.toml:

• mcp>=1.0.0

Add CLI command group (flat style, per CLI AGENTS guidance):

• /Users/joshpurtell/Documents/Github/synth-ai/synth_ai/cli/commands/mcp/__init__.py
• /Users/joshpurtell/Documents/Github/synth-ai/synth_ai/cli/commands/mcp/smr.py

Suggested command:

• synth-ai mcp smr --transport stdio
• synth-ai mcp smr --transport streamable-http --host 0.0.0.0 --port 8765

Rollout Plan

1. Backend contract alignment

• Add project-scoped run aliases in backend.
• Add encrypted_key_b64 compatibility for SMR provider-key set.
• Keep canonical /smr/runs/... routes for backward compatibility.

2. SDK alpha (synth_ai.sdk.managed_research)

• Implement methods listed above.
• Add fallback behavior and unit tests.
• Export from synth_ai.sdk root.

3. MCP alpha (synth-ai mcp smr)

• Implement read-only tools first.
• Add side-effect tools after validation.
• Add tool contract tests (tools/list, tools/call).

4. GA hardening

• Add end-to-end tests against local backend.
• Publish docs and examples.
• Set stability tag policy in docs (Alpha -> Beta -> Stable).

Acceptance Criteria

• Users can control full SMR lifecycle from Python without manual REST wiring.
• MCP tools cover all core operator actions and return consistent typed JSON.
• SDK behavior is stable despite backend route transition (project-scoped + canonical).
• Key upload path is explicit and secure, with deterministic fallback behavior.

Addendum (2026-02-20): Granular usage + dollar-cost semantics

get_run_usage_by_actor(...) now defines two explicit output modes:

• usage_mode="spend_entries": exact cost path sourced from /smr/admin/runs/{run_id}/spend.
• usage_mode="logs_thread_totals": fallback path sourced from run logs token snapshots.

In exact-cost mode, summary and each orchestrator/worker/model/session row includes:

• total_cost_cents, total_cost_usd
• meter_quantities (input, cached input, output, reasoning, and non-token meters when present)
• meter_cost_cents, meter_cost_usd
• token_usage split (input_tokens, cached_input_tokens, output_tokens, reasoning_output_tokens, total_tokens)
• token_cost_cents, token_cost_usd split by:
  • input
  • cached input
  • output
  • reasoning output
• cost_data_available=true

In fallback mode (admin spend unavailable), output includes model attribution + split token quantities but marks:

• cost_data_available=false
• total_cost_cents=None / total_cost_usd=None

Fallback attempts run-level estimated dollars from project usage rollups:

• summary.estimated_total_cost_cents / summary.estimated_total_cost_usd
• summary.estimated_orchestrator_total_cost_cents / summary.estimated_worker_total_cost_cents
• per actor/model/session:
  • estimated_total_cost_cents
  • estimated_total_cost_usd
• summary.estimated_cost_source="project_usage_per_run_token_share"

Allocation rule for estimates: divide run-level project usage cost across actors/models/sessions by token_usage.total_tokens share.

Primary external references

• MCP architecture: https://modelcontextprotocol.io/docs/learn/architecture
• MCP transports: https://modelcontextprotocol.io/docs/concepts/transports
• MCP tools: https://modelcontextprotocol.io/docs/concepts/tools
• MCP Python SDK: https://github.com/modelcontextprotocol/python-sdk
• Synth prompt optimization docs: https://docs.usesynth.ai/prompt-optimization/gepa