Your agent needs a next step. And a reason.

TaskSaber separates deterministic ranking from the model you choose. Over MCP an authorized agent asks for ranked work, reads the reasons behind every score, and holds a task lease while it acts.

Start free Connect a client

Agent access is on. Create a credential in the app under Settings → Agents & usage, then point your MCP client at the endpoint.

Any MCP clientJSON

{
  "mcpServers": {
    "tasksaber": {
      "url": "https://api.tasksaber.com/mcp",
      "headers": {
        "Authorization": "Bearer <your credential>"
      }
    }
  }
}

Claude
Cursor
GitHub Copilot
Google Gemini
Windsurf
Zed
JetBrains
Replit

The credential comes from Settings → Agents & usage in the app. Marks belong to their owners and mean the client speaks MCP, not a partnership.

Ask, claim, report, complete.

This is one agent loop as the server sees it, with the units each call costs, beside what a teammate sees in the app while it happens.

Waiting for the first call…

What your teammate sees

92Fix the double-charge on 3 customer accounts

In Focus

The agent has not asked yet. The task sits where a person would see it first.

Calls: 0
Units: 0
Cost: $0.0000

A full loop is 4 units. The Free plan includes 1,000 a month; beyond any plan’s allowance, units are $0.0002 each and only with a separate opt-in.

Why put an agent behind a ranking.

An agent handed the whole backlog has to guess. This contract hands it one ranked task, the reasons, and a lease. Each of the three properties below is something you can press.

The same answer, every time

Ranking is arithmetic over the task, the clock and your configuration. Ask twice and the agent gets the same task and the same reasons, so its behaviour is explainable after the fact.

next_tasktsk_9f2a · 92

1 answer

Two agents cannot take one task

A claim is a lease with a fencing token. While one agent holds it, a second claim is refused and told when to retry. Writes without the current fence are rejected on the server.

claim_taskgranted · fence 7

1 lease held

Asking is cheaper than browsing

Browsing the backlog costs 2 units a page and hands the model everything. Asking costs 1 unit and hands it one task. The server’s own instructions to the client say so.

Browse: 2 units · 1 page of up to 100
Ask: 1 unit · 1 task

list_tasks ×1

Connect a client.

Create a credentialIn the app: Settings → Agents & usage. It is scoped to one workspace and revocable in a click.
Point your client at the endpointAdd the JSON above to your MCP client with the credential as the bearer token.
Ask for the next taskThe agent gets the same ranked answer you see, with the reasons, and can claim a lease before it works.

15 tools, and what each one costs.

Additional units cost $0.0002 each, capped at $49 per UTC month before subscriptions and taxes. Paid overages need a separate opt-in. Units are not model tokens; what you save depends on your workload and provider.

MCP tools and units per accepted call
Tool	Purpose	Units
`next_task`	Request the next ranked task and its reasons.	1
`explain_task`	Inspect the factors contributing to a task’s score.	1
`list_signal`	Request the ranked signal list.	2
`claim_task`	Request a task lease with a fencing token.	1
`release_task`	Release a held task lease.	0
`report_progress`	Report activity while holding a lease.	1
`complete_task`	Request completion under the current lease.	1
`report_estimate`	Report actual time against an estimate.	0
`find_blockers`	Inspect dependency blockers.	1
`create_task`	Create a new task, subject to the tenant’s active-task quota.	1
`update_task`	Patch a task’s fields with an optimistic version check.	1
`add_dependency`	Declare a blocked-by relation; refused if it would form a cycle.	1
`remove_dependency`	Remove a blocked-by relation; a no-op if it did not exist.	0
`list_tasks`	Browse tasks with filters and bounded cursor pagination.	2
`get_task`	Read one task’s full record and current version.	1

Access is something you grant.

Scoped permissions

Credentials are scoped to one workspace and revocable from Settings in a click. Every tool call is checked against your workspace permissions on the server, and task leases are enforced there too.

Task content can reach your agent

User-selected third-party agents can receive task content. Structured-field redaction does not guarantee privacy for names, financial information or other sensitive material entered in titles and free text.

Check client compatibility

TaskSaber speaks MCP over stdio and HTTP. Cursor connects with the configuration above. Claude Desktop needs a small bridge for remote servers; those instructions arrive with verification. Other MCP clients should work, and we would like to hear if one does not.

The contract, machine-readable.

Everything on this page is generated from one document, served at /.well-known/tasksaber.json. An agent or a build can read it directly; the hash changes when the catalog does.

What it contains

Release status and which surfaces are verified live: sales, the agent endpoint, each provider, iOS and Android.
The plan catalog: limits, features and trial terms, with hash 518300882e9f….
The 15 tools with units per call, unit price and the monthly cap.
The interface declaration: protocol mcp, and whether the deployment is verified.

Open the contract View plans

GET https://tasksaber.com/.well-known/tasksaber.json

{
  "name": "TaskSaber",
  "status": "owner-configured-release",
  "availability": {
    "salesEnabled": true,
    "agentEndpointVerified": true,
    "verifiedProviders": ["github","jira","slack"]
  },
  "catalogHash": "518300882e9f7fb0dec2e019610db24aac7faff493580ee25e10dae1fc90cb36",
  "interface": { "protocol": "mcp" },
  "pricing": {
    "model": "metered-per-decision",
    "unitPriceUsd": 0.0002,
    "monthlyCapUsd": 49
  },
  "catalog": { "agents": { "tools": [ …15 tools… ] }, … }
}