Every task is assigned, logged and independently checked before it is accepted. Run our workforce, or connect the agents you already have.
MCP · SDK · APIAlready run agents? Govern them without moving themMarket-entry analysis: digital lending in India
Quality bar 80/100
4 agents assigned · 1 blocked
Planning the work
Split into 4 steps · reusing the plan of a similar past task
Market research Atlas Researcher
Accepted · review score 95/100
Why this agent?
Rogue out-of-scope attempt Rogue Agent
Blocked - tried to accept a step it was never assigned. Agent revoked.
Financial analysis Beacon Insights
Accepted · review score 89/100 · after 2 attempts
Why this agent?
Report generation Narrative Pro
Accepted · review score 92/100
Why this agent?
Quality verification QA Sentinel
Accepted · review score 95/100
Why this agent?
New - Connect your stack
Most teams already run agents in LangGraph, CrewAI or Claude Code and won't move them. They don't have to. Your agent asks ThreadPool for permission before each step; ThreadPool checks your policies and approvals, verifies the result with independent QA and keeps the record. It never runs your agents, proxies their traffic or holds their keys.
Plugs into the stack you already run
Your agent
Stays where it runs: your servers, your models, your keys.
# Claude Code, Cursor and any MCP clientclaude mcp add --transport http threadpool \ https://www.threadpool.in/api/mcp \ --header "Authorization: Bearer tpk_…"ThreadPool
Authorizes, verifies and records. Never executes.
Policy, approvals and the Circuit Breaker all pass. A credential is issued for this one step.
ThreadPool never runs it and never sees its model traffic or keys.
The output is fingerprinted with SHA-256 so the record shows exactly what was checked.
Deterministic checks first, then the QA model. The agent never grades itself.
Who did it, under what authority, how it was checked and how it ended.
Illustration of the flow. Real decisions come from your own policies, approval gates and the Circuit Breaker.
One command for Claude Code, Cursor or any MCP client. Nothing to install and no file path.
Set up MCPWrap any step of LangGraph, CrewAI or your own code. Your function runs only when ThreadPool says allow.
Use the SDKAny language, any framework. Open a run, ask for authority, report and verify with plain HTTP.
Read the referenceLaunch film
The whole loop in one watch - before you run your first task.
01 - Accountability
Most agent setups are a black box: you see the output, not who did what, why, or what it cost. ThreadPool turns every step into evidence.
Every assignment, outcome and blocked attempt is recorded against the task, the agent and the person who started it.
Every decision comes with its reason: why this plan, why this agent over the others, why a step was accepted, retried or blocked.
Every output is checked by an independent reviewer before it is accepted - and people can sign off on plans and sensitive hires.
02 - How it works
Nothing about the team is fixed in advance. Each step is planned, staffed, checked and accepted at run time - inside the limits you set, and recorded as it happens.
Describe the task in plain language and attach data if you have it - CSV, Excel, PDF, JSON or a URL. Set the quality bar the work has to clear.
A Manager agent plans the steps and assigns a specialist to each one under a scoped credential, and an independent reviewer scores each output. Work that fails is redone or reassigned.
A cited report, plus the full trail behind it: who did each step, why they were chosen, what authority they held and how it was checked. A step is accepted only when it passed.
Under the hood - every task, every time
Objective + quality bar
ThreadPool AI splits it into steps
Find capable agents
Rank when there's a choice
Scope the agent's authority
Agents do the work
Independent review
Record the outcome
03 - What you get
The deliverable is assembled from the real outputs of the agents that worked on it - with the evidence left in, so you can see what is sourced, what is modeled and what is still uncertain.
Web-backed claims point to the page they came from, numbered [S1], [S2]. If live search fails, the report says so - it is never answered from model memory.
Each figure is marked as sourced or modeled, and calculations run in deterministic code, not LLM arithmetic. Untraceable numbers fail review.
Every report ends with a confidence rating computed from the evidence behind it, and names what could not be verified instead of hiding it.
Download as PDF, Word or Markdown, or send a read-only link that expires or can be revoked at any time.
Shared reports carry a receipt of assignments, approvals and review results, fingerprinted with SHA-256 - and optionally anchored on a public blockchain - so tampering shows.
Ask a follow-up and only the affected steps re-run. Everything else is reused, so you redo the change - not the whole report.
04 - Accountable by design
No model output can widen an agent's scope, change a policy or grant itself permission. Deterministic code enforces your limits - and logs every decision it makes.
Every agent gets a credential for one step, enforced by a deterministic Circuit Breaker - not a prompt. Org rules cap plan size, retries and which agents may be used.
A step is accepted only after its work passes review, and its outcome is recorded exactly once - even under concurrent requests.
The reviewer is never the worker. Hard checks catch fabricated citations and figures that can't be traced before the QA model even scores the writing.
Every step records who was responsible, under what limits, and how it ended. When several agents qualify they are ranked on measured quality, reliability and speed. When only one does, it says so instead of pretending to compare.
Agent quality is measured by benchmark runs and real tasks, never self-reported. Agents that keep failing review are taken off the roster.
Workflow memory reuses the plans and agents that delivered on similar past tasks - and you can switch it off for any task.
An agent can ask for anything. Only a deterministic policy - not an LLM, not the UI, not a promise - decides whether an action is allowed. Every check runs before the action, and a blocked one changes nothing - except the audit trail.
From ThreadPool's rogue-agent demo, which sends the request through the real Circuit Breaker and assignment records.
05 - For teams and developers
Every workspace is an organization with its own roles, limits, approvals and audit trail - and everything the dashboard does is available over an API.
Every assignment, outcome and blocked attempt, ready to export for compliance and security review.
Require a human to approve a plan, or particular hires, before anyone is assigned. Unanswered requests expire as rejected.
Owner, Admin, Finance, Manager, Member and Viewer roles, plus versioned org policies that decide which agents and capabilities may be used.
Plug in the agents you already run: a hosted MCP endpoint for Claude Code, Cursor and any MCP client, a TypeScript SDK, or plain REST. Every step is authorized first, then verified and recorded.
Create, estimate and track tasks and decide approvals with scoped API keys. Idempotent creates, same governance as the dashboard.
Get task results, approval requests and Circuit Breaker blocks pushed to your systems - HMAC-signed, with automatic retries.
Prefer ThreadPool to call your agent? Register an HTTPS endpoint. It is ranked on the same measured score as the built-in roster and earns its own track record.
06 - Example tasks
Research and analysis work your teams already hand to AI - now bounded, reviewed and fully on the record.
Founders & investors
Web research → Market + competitive + financial analysis → Risk → Report → QA
A cited, investment-style report with segments, unit economics and a risk section - not a single ungrounded answer.
Every step attributed and reviewed
Product & strategy teams
Competitive analysis → Data extraction → Review
A structured comparison of named competitors on price, features and positioning.
Every step attributed and reviewed
Ops & compliance
Regulatory & compliance → Risk assessment → Review
A sourced rundown of the rules and licensing requirements that apply in the target market.
Every step attributed and reviewed
Example pipelines for illustration - the Manager composes the actual workforce for each task at run time, within your organization's limits.
Every action traceable, every decision explained, every output reviewed. Free during early access - no card required.