Mantri

Hosted execution plane for coding agents

One control plane for every coding agent.

Use the AI subscriptions and machines you already pay for. Mantri runs each task in a sandbox and opens a PR after you approve.

Create an account

Sign in to Mantri

Use Google, GitHub, or a one-time code sent by email.

Continue with Google Continue with GitHub
or

New to Mantri? Create an account

  1. 01Intake
  2. 02Placed
  3. 03Run
  4. 04Verified
  5. 05Approve
  6. 06PR
The board: paper colour shows the agent. Each task moves right until you approve it.

SOC 2 ready.

Credentials never leave your hosts.

Branches only, never merges.

Sandboxes are destroyed after each task.

Features

Send work in. Get reviewed pull requests out.

These parts of Mantri work today. Open a group to see its features and screens.

Mission control: tasks handed back, pass rate, time to pull request, service status per agent, and items that need you.
Mission control. Weekly figures, agent health, and the items waiting for you.
01 · Submit work

Turn a sentence or an issue into a bounded task.

5 features
  • New task composerWrite the intent, pick the repository and gates, and see time and cost before you dispatch.
  • GitHub issue intakePaste an issue URL. Mantri snapshots the title and body so later edits cannot change the task.
  • Rachna proposalsApproved Rachna proposals arrive as tasks with their graph context. Optional.
  • Acceptance gatesEach project lists the checks every task must pass, such as tests, typecheck, and architecture rules.
  • Estimate before dispatchShow expected time and credits from similar past tasks, and reserve credits up front.
New task dialog with intent text, repository, acceptance gates, who builds it, and a time and cost estimate.
New task. Intent, repository, gates, and an estimate before dispatch.
02 · Fleet and capacity

Pool your machines, your cloud, and managed capacity.

5 features
  • Self-hosted runnersInstall the runner CLI on any Linux or macOS machine, connect it, and update it safely.
  • Your cloud runnersSpawn runners in your own GCP project with Workload Identity. You pay Google directly.
  • Managed burstBorrow Mantri-managed VMs when your own capacity is full. Paid with credits.
  • Account discovery and headroomThe runner reports which agent accounts exist and how much quota each has left. Keys never leave the host.
  • Drain, spawn, and orphan cleanupDrain a machine safely, spawn cloud runners, and find and delete forgotten VMs.
Capacity screen listing machines, agent accounts with quota left, cloud VMs, and the sandbox contract.
Capacity. Machines, agent accounts, and cloud runners in one list.
03 · Agents and models

Drive many coding agents through one protocol.

2 features
  • Agent catalog over ACPDrive Claude Code, Codex, Gemini CLI, Copilot CLI, Goose, and OpenHands through one open protocol.
  • Pick agent and model per taskChoose an agent, model, and machine yourself when you do not want automatic routing.
Agents and models screen comparing pass rate and cost per task for each agent.
Agents and models. Pass rate and cost for each agent you connect.
04 · Routing and policy

Rules you can read decide who builds what.

3 features
  • Placement explanationsEvery placement shows the chosen route and why each other candidate was rejected.
  • LimitsCaps per task, per person, and per day for time, turns, concurrency, and credits.
  • Autonomy levelsPlan only, Verify, or Auto PR. Merge and force-push never exist.
Routing and policy screen with autonomy levels, ordered routing rules, limits, and the last placement decision.
Routing and policy. Autonomy level, rules in order, and limits.
05 · Safe execution

Each task runs in a fresh sandbox and is verified.

5 features
  • Disposable sandboxEach task runs in a fresh worktree with no network beyond an allowlist. It is destroyed after.
  • Fenced leases and retriesOnly the owning runner can report results. Retries create new attempts and keep history.
  • Independent verifierA separate check confirms the result. If it fails or is unsure, the task does not reach review.
  • Mantri-made commitsMantri creates the commit from the worktree. Agents never hold GitHub write access.
  • Cancel and recoverCancel any task. After a crash, Mantri recovers and settles credits once.
Task screen with the journey from queued to pull request, the agent plan, budget, acceptance gates, and cancel and recover.
Task. Journey, plan, budget, gates, and recovery for one task.
06 · Review and hand-back

A person approves each hand-back. Mantri never merges.

4 features
  • Review queueOnly verified work appears here, with risk, gates, cost, and age.
  • Approval with evidenceThe dialog pins the commit, evidence digest, gates, risk, and target branch.
  • Approve with editsEdit the patch or PR text. Mantri reruns every gate before hand-back.
  • Branch and PR hand-backPush to a scoped branch and open one PR, even after retries.
Review queue with verified tasks, their risk, gates, cost, and age.
Review queue. Only verified work reaches it.
Approval dialog that pins gates, verifier result, changed files, cost, the diff, and the pull request title.
Approval. The dialog pins the evidence you approve.
07 · Observability

See each run and decision in one place.

2 features
  • Live boardFive lanes from Queued to Ready for review, with progress and agent on each card.
  • Event stream and audit trailEvery step, decision, and denial in an append-only, redacted log.
Network board: each agent is a line and each task a train between the stations queued, placing, running, verifying, awaiting you, and handed back.
Live board. Each agent is a line; each task moves between stations.
Runs and health screen with a trace of one task, its spans, and the event stream.
Runs and health. A trace for each task and the event stream.
08 · Spend and enterprise

Prepaid credits with a receipt for each change.

2 features
  • Credit ledgerPrefunded credits with reservations, use, and releases. Every change has a receipt.
  • Organisations and rolesOwners, admins, and members, with invitations and an organisation switcher.
Spend and budgets screen with credit balance, spend by provider, team budgets, and the credit ledger.
Spend. Credit balance, budgets, and the ledger.
Organisation screen with members and roles, integrations, and the audit log.
Organisation. Members, roles, and integrations.