Development — What exists today, what is limited

Status

A clear view of the current stage.

A mechanism existing is not a statement that every workflow built on it is supported or proven.

exists

Mechanisms in place

Repository evidence with source references and provenance, from a bounded census that records where knowledge was not collected
Evidence served through local MCP tools: exact lookup, listing, what depends on a file, existence checks that answer known or unknown and suggest the nearest real names, and patterns your team has approved
Tools offered only where the repository’s knowledge can answer them, with held-back tools named; a stale-knowledge banner when the code has moved past the last build
Hook-delivered operating guidance and project rules, re-delivered after /clear and compaction
Intake that records answers locally and logs decisions with who made them and when, plus an optional note on why, inside a task with written scope and acceptance criteria; readiness decided by a deterministic check
Configured gates for research prerequisites, Intake readiness and edit scope, each naming the missing condition and a next action
Context packs sealed against their source records and checked for staleness; development checkpoints with a mandatory record of deviations from the plan
AtlasBound’s own Handoff skill, shipped with the package and enabled with a one-line opt-in, with checkpoint-backed context packs for active Intake work and a narrative fallback without Intake
Repository records authored by an agent pass deterministic validation and a separate verifier before they count as verified; missing proof withholds the claim

limits

Known limits

Evidence covers mapped ground only. The rest is named unknown, and a human answer applies within its recorded task, not to the whole project.
Effectiveness across real teams has not been established or published. No percentage or improvement figure is claimed.
Which workflows can be relied on, and under what conditions, is still being qualified.
AtlasBound runs with Claude Code today. Codex support is in development and not yet supported; other clients are not confirmed.
Advisory context can be ignored by the agent. Only a gate can block an action, and hooks fail open on error, so silence is never treated as approval.
Current research-gate profiles cover test changes and frontend component changes. Backend and database profiles are not qualified. The gate covers the editor’s direct write tools, not shell writes. A research record shows the agent consulted the evidence, not that the change is correct.
The research gate has been qualified on test fixtures; its behaviour on customer repositories is part of what early-access engagements establish.
Reference checks ask for confirmation rather than block in the current release. Blocking on unknown references is reserved for human-sealed, measured patterns and is not enabled.
Project rules and conventions exist only where they have been defined and are delivered as guidance; the edit gate asks about departures, it does not refuse them.
AtlasBound does not collect runtime data itself. Intake records observations supplied through tools you connect and authorise.
Context-pack seals establish source freshness, not the correctness of every claim or whether the agent read and understood the pack.
Records are local and single-user: they sit on the machine where the work happened, and sharing them across a team is up to you. Retention and backup are your practices; team governance, access control and a hosted knowledge store are not supplied.
A decision’s recorded author is the name entered with it, not an authenticated identity, and local records are not a tamper-evident audit trail.
By default Intake asks the developer running the session. Sending a question to someone else needs the optional decision exchange, which runs through AtlasBound’s hosted service.
Hooks and gates are workflow controls for a cooperating agent, not an adversarial security sandbox.

direction

Where it is heading

Preparing early access for engineering teams on large or multi-team codebases
Qualifying workflows and naming supported environments once verified
Narrowing the remaining gaps in repository extraction, with unsupported ground kept visible
Growing project rules and verified conventions from repeated mistakes
Completing Codex support, announced as supported only after verification

Plain claims

What AtlasBound does not claim.

Whole-repository understanding: Claims are tied to evidence for the ground that was mapped, nothing wider.
Universal or production-ready behaviour: It is not described as ready for every stack, team or environment.
Perfect reliability: It does not promise the end of hallucinations or of consistency drift. It adds mechanisms designed to make them easier to see and less likely to repeat.
That the agent always follows the context: Whether an agent read and applied a context pack is not measured or scored. Gates, not context, are what can stop an action.
Support for every agent: AtlasBound runs with Claude Code today. Codex support is in development; anything else is announced only after it is verified.

Early access

What we establish with your team.

AtlasBound is offered to engineering teams through direct contact while it is in active development. There is no self-serve signup yet. Each engagement is scoped with the team before it starts.

What is mapped: Which repository facts are mapped for your codebase, and where the unknowns are.
What Intake surfaces: Whether Intake raises the questions your team would otherwise settle late, in review.
What you protect: Which of the supported surfaces you want gated (today: test changes, frontend component changes and Intake readiness), and how the gate’s reasons and next actions read to your developers.
What stays with you: Repository records, Intake history and context packs stay in your environment; the optional decision exchange stays off unless you choose to turn it on.

FAQ

Questions worth answering plainly.

Is this generally available?

Not as a self-serve product. AtlasBound is offered to engineering teams through direct contact while it is in active development. An earlier build of the local package is on npm; it predates much of what this site describes and is not the supported build.

What does “Reinforced” mean?

The environment repeatedly supplies and applies evidence, human decisions, verification and project rules around the agent. It does not mean reinforcement learning or model training.

Does it replace our coding agent, IDE or rules file?

No. Keep your agent and your CLAUDE.md. A rules file is advisory text in context. AtlasBound adds source-linked repository records retrieved on demand, decisions recorded with who made them and when, freshness-checked context packs, and configured gates that can block a protected edit.

Where are records stored?

Locally, in your environment, under your storage and backup practices. AtlasBound does not operate a hosted store for your repository knowledge, Intake history or context packs. The one exception is the optional decision exchange, which is off by default. When you turn it on, the exact question you preview and approve goes to a named teammate through AtlasBound’s hosted service, which keeps that question and its answer. Your coding agent’s own provider terms govern what the agent sends to its model while it works.

What can AtlasBound block today?

Edits to surfaces you protect when the required research or a ready Intake is missing, and edits to environment files. Research requirements currently cover test changes and frontend component changes. Where a reference check is set up, an unknown reference leads to a confirmation request, not a block. See Product and the limits above.

Do developers have to change how they work?

They keep their agent. Intake asks people only what research cannot settle, and gates apply only to the surfaces you protect. Everything else proceeds as normal.

Is the workflow example a real run?

No. The workflow example is an illustration with fictional names. It is not a screenshot or a recorded result.

Does it work with agents other than Claude Code?

AtlasBound runs with Claude Code today. Codex support is in development; it will be announced as supported once verified. Other clients are not confirmed.

Where do I ask a question?

Email [email protected]. See Contact.