Control plane for coding agents

Give coding agents a controlled path from issue to pull request.

Terrarium helps engineering teams publish focused agents, send them work from Linear or GitHub, run every task in an isolated sandbox, and follow each run and its outputs through review.

Open Terrarium See how it works

For engineering teams working in Linear and GitHub.

From delegation to review

Move agent work forward without making it invisible.

A coding agent can move quickly. Your team still needs to decide which instructions it follows, which repository it can access, and how the result returns to a human. Terrarium keeps those decisions attached to the run.

Define the role once

Set the agent's instructions, model, capabilities, and repository access, then publish a version that new runs can use.

Published agent profiles: Configure a focused role and publish an immutable version of its instructions, model, capabilities, and repository access.
Credential separation: Vercel Connect holds GitHub and Linear credentials outside agent prompts and repository sandboxes.

Start where the work lives

Launch a task in Terrarium or route work from Linear and GitHub to a published agent.

Linear and GitHub intake

Terrarium
Linear Agent Sessions
GitHub conversations

Published agent profiles

Send it a scoped task

Isolated execution

Dedicated session and sandbox

Linear and GitHub intake: Route work from Linear Agent Sessions and GitHub conversations, or launch a run directly from Terrarium.
Isolated execution: Each task runs in its own session and Vercel Sandbox rather than sharing conversation state or a working filesystem with another run.

Review one connected record

See the selected agent version, execution timeline, available context snapshots, follow-up input, and linked outputs in the run.

Published agent version
Execution timeline
Available context snapshots
Follow-up input
Linked outputs

Human checkpoints: Agent-initiated Linear changes require user approval, and waiting runs can accept follow-up input before work continues.
Run visibility and outputs: Inspect status, timeline events, redacted prompt and context snapshots, repository details, and linked artifacts such as commits and pull requests.

How it works

A clear route from task to handoff.

Terrarium connects agent configuration, task intake, contained execution, and review in one workflow.

Illustrative example

Configure
Connect
Run
Supervise
Ship

01 / Configure
Publish a focused agent
Define its role, operating instructions, model, capabilities, and allowed repositories. Publishing creates an immutable version for new runs.
Your action
Choose the role, instructions, capabilities, and repository access.
Terrarium action
Terrarium publishes the configuration as an immutable version.
Evidence or control
Frontend maintainer v4 records the configuration and allowed repository used by new runs.
Agent profile
Example
Frontend maintainer
@terrarium frontend-maintainer
Published v4
Role
Frontend implementation and accessibility
Capabilities
Session identity · Coding workflow
Repository access
example/terrarium-webWrites enabled
main · default target
Illustrative agent configuration for Frontend maintainer v4, published with example/terrarium-web as its allowed repository.
02 / Connect
Send it a scoped task
Launch a run in Terrarium, delegate a Linear issue, or mention Terrarium in GitHub. The task stays connected to its source.
Your action
Launch in Terrarium or route work from Linear or GitHub.
Terrarium action
Terrarium routes the source context to the selected published agent.
Evidence or control
EXAMPLE-142 stays linked to Frontend maintainer v4 and example/terrarium-web.
Task intake
Example
LinearConnected
GitHubConnected
LinearEXAMPLE-142
Improve run timeline
Published agent
Frontend maintainer v4
Repository
example/terrarium-web
Illustrative intake keeps EXAMPLE-142, Frontend maintainer v4, and example/terrarium-web connected in the same task handoff.
03 / Run
Run in isolation
Terrarium starts a dedicated session and sandbox, prepares the selected repository and branch, and applies its configured access.
Your action
Send the scoped task to the published agent.
Terrarium action
Terrarium prepares an isolated session and sandbox for the task.
Evidence or control
run_example_142 records the dedicated execution environment, repository, and branch.
Run environment
Example
EXAMPLE-142 · Frontend maintainer
run_example_142
Running
1. Dedicated session started10:42:08
2. Sandbox created10:42:11
3. terrarium/example-142-run-timeline10:42:16
4. Repository context prepared10:42:19
Illustrative execution record for EXAMPLE-142, with a dedicated session and sandbox targeting terrarium/example-142-run-timeline.
04 / Supervise
Follow the run
Watch status and timeline events and provide input when a run waits. Agent-initiated Linear changes require user approval.
Your action
Review the waiting point, then approve or provide follow-up context.
Terrarium action
Terrarium shows when execution is active, waiting, completed, or failed.
Evidence or control
run_example_142 shows the waiting point and the proposed Linear change before approval.
Run timeline
Example
Frontend maintainer v4
EXAMPLE-142 · example/terrarium-web
Waiting
Linear change requires approval
Move EXAMPLE-142 to In Progress and add a progress update.
ApproveDecline
Follow-up input remains attached to this run.
Illustrative checkpoint for run_example_142: the proposed Linear update remains visible while the run waits for approval.
05 / Ship
Review the handoff
Open linked outputs—including a pull request when the run produces one—and review the result in the source-control workflow.
Your action
Open the linked output and continue human review.
Terrarium action
Terrarium attaches available commits, pull requests, and provider links to the run.
Evidence or control
This illustrative run links terrarium/example-142-run-timeline, a commit, and pull request #284 in one record.
Run outputs
Example
Improve run timeline
EXAMPLE-142 · run_example_142
Completed
Branch
terrarium/example-142-run-timeline
Commit
8f42c1d9a706
Pull request
#284 · Improve run timeline
The output is linked for human review; merge, deployment, and release stay outside this handoff.
Illustrative handoff for run_example_142. Pull request #284 is shown as a linked output this example run produced, not a guaranteed result.

Current product boundary

Available now

A run can expose status, timeline events, follow-up input, context snapshots, and linked outputs such as commits and pull requests when it produces them.

Outside the current boundary

A pull request is a possible output, not a guaranteed result. Terrarium does not automatically merge, deploy, or release code.

Capabilities

The context agents need. The visibility teams need.

Technical boundaries support the workflow instead of becoming the workflow.

Controls

Control the run without sitting inside it.

Terrarium keeps identity, repository access, execution state, and credentials in explicit product boundaries.

Credentials: Provider-owned
Session and filesystem: Isolated by task
Agent version: Immutable at run start
Linear mutations: User approval required

Bounded repository access

Attach only approved repositories to an agent, choose a default branch, and configure whether repository write actions are allowed.

Traceable execution context

The run keeps its published agent version attached. Manual and Linear attempts also capture a redacted snapshot of the prompt and context handed to the executor.

Provider-owned credentials

GitHub and Linear grants remain in Vercel Connect. Terrarium stores organization-scoped installation references instead of provider OAuth tokens.

Visible waiting points

Runs show when execution is active, waiting, completed, or failed, so an operator can provide follow-up context or review the result.

Put agents to work

Bring focused coding agents into your engineering workflow.

Publish the agent, choose the repository, and follow the run from request to review in one place.