Tembo

Executive Summary

Disclosure: Ry Walker is Tembo’s co-founder and CEO. This is a founder-authored profile; product documentation and company claims are distinguished from external experience below. ^[1]

Tembo is a coding-agent execution and orchestration platform. It connects agents to repositories, dependencies, tickets, and tools, then gives teams a shared place to observe work and review changes. The current product spans managed cloud execution, automations, pull-request review, and self-hosted deployment. ^[2]

The practical reason to evaluate it is moving recurring or parallel engineering work off individual developer machines while keeping the team's chosen harnesses. It does not eliminate the need for acceptance criteria, reliable tests, or human judgment about generated changes.

See Cloud Coding Agent Platforms for the broader category.

Product and Workflow

Tembo documents Claude Code, Codex, Cursor, OpenCode, Amp, Pi, and fx as available harnesses. Teams choose a harness and model for an agent, with session-level overrides. This is a supported catalog, not a promise that arbitrary agents or every model work interchangeably. ^[6]

Typical inputs include GitHub activity, Linear issues, Sentry errors, and Slack messages. MCP connections extend what agents can read and change. That breadth also makes integration permissions part of the evaluation: a code-review process alone does not govern every action a connected tool might expose. ^[6][8]

Recent releases added a full PR review workspace in July, private versus workspace-visible sessions in August, and broader public API resources, native planning mode, and workspace-level subscriptions in September. Keeping project environments ready reduces setup waits but incurs compute usage; this review did not benchmark latency. ^[5]

Deployment and Data Boundaries

Managed cloud

The sandbox documentation describes a dedicated Linux VM per session. It also qualifies ephemeral execution: prepared projects retain cached repositories and dependencies until rebuilt or removed. Do not read “ephemeral sandbox” as “no code is ever stored.” Model requests have a separate boundary: provider retention follows the selected service or BYOK account terms. ^[9]

Projects also preserve the output of setup scripts in their prepared environment. Changes to dependencies, setup commands, or build-time variables require rebuilding that environment. This makes project configuration a maintained build artifact, rather than an automatic guarantee that every session starts from the newest setup. ^[7]

Self-hosted

The documented stack uses a NixOS machine image with the web application, API, agent workers, scheduler, and supporting services. Updates are pulled using a license key; the operator controls networking, secrets, backups, and daily operations. Custom model endpoints are supported when they implement the expected APIs. ^[10]

AWS and Azure guides describe vendor-provided images and setup flows; AWS hardware acceleration depends on suitable bare-metal or nested-virtualization support. ^[4][11] A detailed GCP guide is published, but the overview still labels GCP and Kubernetes support early access. Confirm the supported deployment and upgrade path for a purchase rather than assuming every advertised cloud has the same maturity. ^[12][10]

Self-hosting locates application data and runtime configuration in the customer's environment. It does not, by itself, make external model endpoints or connected services local. The documented update path also uses outbound HTTPS. Fully disconnected operation needs an explicit deployment and inference design. ^[10]

Security evidence

Tembo's security page describes SSO, role controls, audit trails, encryption, and retention settings. Its trust center lists SOC 2 and ISO artifacts behind an access request. These are vendor-published controls and report listings; this refresh did not inspect the restricted audit reports or test isolation. Verify scope, covered services, and deployment-specific responsibilities during procurement. ^[13][14]

Pricing and Licensing

Current public USD plan pricing is: ^[3]

The Free plan card says one user, while the same page's comparison table and billing documentation say up to three. Billing documentation also includes one Free project, while the Projects guide still says paid-only. Confirm these entitlements in the workspace; the published pages are inconsistent. ^[3][15][7]

The old per-task credit model is stale. Gateway token usage and VM compute draw from one dollar balance. Monthly allowance expires; prepaid funds do not. Pro and Max support capped overages. On Free, exhausted funds block new sessions and running sessions are checked after model calls. ^[15]

Compute is billed at $0.0403 per vCPU-hour plus $0.0130 per GiB-RAM-hour, measured by running seconds. For example, the listed 4-vCPU/8-GiB Medium VM is $0.2652/hour before inference. Non-running VMs do not accrue compute charges. This is a rate example, not a workload cost estimate. ^[15]

Standard BYOK and supported connected subscriptions work on Free; AWS Bedrock and Vertex AI keys require a paid plan. BYOK inference is charged through the provider account instead of Tembo, while Tembo compute still consumes allowance. Supported OpenAI and Anthropic access depends on the harness and workspace configuration; subscription limits and provider terms still apply. ^[16]

Commercial self-hosting requires a license. The separate Agent Studio project is MIT-licensed and stores agent definitions in Git; its hosted Tembo integration is optional. Its open-source license should not be generalized to the entire Tembo platform. ^[4][17]

What Developers Say

A first-hand public account exists: Anurag Dhungana's November 7, 2025 review describes moving work to Tembo because parallel local agents overwhelmed his laptop. He reports receiving PRs for a writing-app redesign, using agents to update his personal site, and experimenting with scheduled MCP workflows. He also reports failures of continuity that led to incorrect output. ^[18]

This is one developer's historical experience, not a current reliability benchmark or enterprise reference. His introductory credit amount predates today's pricing and is not used above. The previous profile's statement that no substantive first-hand product review was found was too broad.

Company customer logos, quotes, and activity figures can identify reference leads; they do not establish comparative completion rates or cost savings. This review does not infer market share or measured superiority from them.

Company Continuity

In his April 9, 2026 retrospective, Walker confirms that the company shut down its earlier Tembo Cloud service and rebuilt around autonomous software maintenance. This is historical context for the company, not the description of today's coding-agent platform. The source explains the orchestration thesis from the founder's perspective; it is not independent validation of that thesis. ^[19]

The published funding total establishes reported financing, not current cash, profitability, or a support guarantee. A team making the platform operationally important should establish export, continuity, and support expectations in its agreement. That evaluation remains relevant whether using managed cloud or maintaining a licensed deployment.

Assessment

Best fit: teams that want chosen coding harnesses to work against connected tickets, alerts, and repositories, with a common execution and review surface. Evaluate a representative recurring workflow, including environment preparation, connected-tool permissions, failure recovery, review effort, and total compute plus inference cost.

Less compelling: a developer whose local harness already handles the required work, or a team that gains little from shared integrations and unattended execution. Choosing a platform is justified by the coordination work it removes, not simply by adding another interface to the same model.

The current evidence supports a broader orchestration platform than the June snapshot described: more harnesses, review and planning workflows, dollar-based billing, and concrete self-hosting guides. Remaining evaluation work includes deployment-specific controls, documentation inconsistencies, and actual task economics. Sources were checked September 15, 2026; this refresh did not run a product trial, inspect restricted audits, or benchmark competing platforms. See the research methodology.