Status and limits
What works in version 0, how it is tested, and what is missing.
Writ is at version 0. You can plan, approve, and apply OpenTofu stacks from a Git repository with it. It has policy, drift checks, imports, and one mask over the plans, run output, and state index rows it shows.
Here's what's built, how it's tested, and what's missing, so you can judge what to trust it with. Our advice today: don't move production infrastructure to Writ.
What you can use today
- A control plane,
writ serve, for OpenTofu 1.13.9 stacks in a Git repository: a plan of each commit you or your CI job sends it, approvals bound to one exact plan, and an apply of the saved plan that was approved. See How the control plane works and Approvals. - State on the
httpbackend thatwrit serveprovides, sealed and versioned, or on a stack's ownbackend "s3", where a plan runs with a plan role that can't write the state. - One mask over everything
writ servestores or shows, checked byjust verify-mask. See Secrets and the mask. - A private filesystem view and its own provider uid for each
tofuprocess on a team's control plane, and an allow-list of module sources. - OPA policy, checked at the plan and again before the apply, with the same verdict as
conftest. See Policy. - Drift checks on a schedule, the state index of what each stack manages, and
writ state adoptfor a state thatwrit servedidn't write. writ import bulk, for many existing objects into an OpenTofu stack, with references between them and afor_eachfor objects of one shape, andwrit import acceptof the plan that imports them.writ mcpand the writ skill, so your agent plans a commit, reads the masked plan and its policy verdict, and imports objects. See Use Writ with an agent.--jsonon every command butserve,mcp,ci, andhelp.
For every command, see the commands page.
How it's tested
You don't have to take the list above on faith. Writ's end-to-end tests build writ and run it against real OpenTofu 1.13.9 and real provider binaries, offline once the providers are cached:
hashicorp/random,hashicorp/tls, andhashicorp/local, throughwrit serveon itshttpbackend.hashicorp/randomand the built-interraform_data, throughwrit serveon a stack'sbackend "s3", against an S3 stand-in with per-key policies.hashicorp/nomad, against a local dev agent, forwrit import bulk.hashicorp/aws, offline, against a stand-in and against AWS Emulator, forwrit import bulkandwrit import accept.
None of these tests reaches a real cloud host.
What you'll run into today
- There's no proof in a real cloud account yet. S3 state, the credential broker's roles, and imports run against stand-ins only.
- You get one node. A second
writ serveon the same database fails the other's runs. - Many client addresses together can fill the 1,000 places of
writ loginsign-ins that wait, and lock everyone out of signing in for up to 10 minutes. One address can hold only 10. See Sign-ins that wait. - Nobody has run the GitLab CI component on a real GitLab yet, and there's no GitHub flow.
- You won't find notifications, a test command, editor support, or cost estimates.
writ import bulkcovers the types it was measured on; see Import many objects.
Not built yet
These are missing today. If you need one of them, plan around its absence.
writ serve
writ login --print-tokenandwrit serve --init. A sign-in through a real GitLab, in either flow. That GitLab takeswrit login's loopback redirect on any port is inferred from the Doorkeeper version it pins, as Set up writ login explains.- The GitHub Actions token request in the CLI. In GitHub Actions, your job sets
WRIT_ID_TOKENitself. - Roles for clouds other than AWS. For another cloud, your provider gets its credentials from
provider_env. - Plans of several stacks and
--changedon the server. - Under
writ serve --dev: the issuers, teams, grants, approval rules, and environments of a full config.--configtakes onlyforge,repo,tofu, andpolicyblocks andpoll_every. GET /v1/plans/{id}/events,POST /v1/lookups, alerts, among them alerts of drift, and merge request notes.- Console to code: code and a merge request for an object created in the console.
- A second node. Postgres makes one possible later. Sign-ins of
writ loginthat wait for their person live in one process's memory. If you run twowrit serveprocesses on one database, they fail each other's work. At its start, each one fails the plans and runs the other is making, then every minute after it fails the runs the other applies, and releases their state locks.
Approvals and applies
- Run events and cancel:
GET /v1/runs/{id}/events, run events streamed aswrit.event/v0, andPOST /v1/runs/{id}/cancel. Your client polls the run. - Withdrawing, applying, and cancelling from the control UI. You can only approve there. The control UI's audit page waits for
GET /v1/audit. - Merge request notes, and plans of several stacks as one plan.
- Destroy through the control plane, and the GitHub Actions jobs.
- A second factor before an apply. Writ would check it in the transaction that moves a plan to
applying.
Repositories
- GitHub, with a GitHub App.
- Plans of several stacks,
base, and the stacks a merge request changed. - Merge request notes, commit statuses, and the write token.
- A webhook hint that makes a poll happen sooner.
- Discovery of a stack's source digest. Plans record their own.
- Removing a repository through the API. If you use a pinned list, it removes rows when your config drops them.
- Garbage collection of blobs that no commit refers to.
Policy
- A policy that requires more approvals.
- An override of a deny.
- Policy read from a repository that an admin activates. Today policy comes only from the server's config.
writ policy testandwrit policy eval. For now, useconftest verifyandconftest test.
This page is built from docs/status.md in the repository. It is also served as plain Markdown.
Writ compared with other control planes
Writ against Atlantis, OTF, Terrakube, Spacelift, HCP Terraform, env zero, and Scalr, from each tool’s own pages, with sources and dates, and what they have that Writ doesn’t.
How the control plane works
How writ serve plans a commit, runs OpenTofu apart from its keys, keeps state, checks drift, and decides who can do what.