Workspace Atlas Project context infrastructure
GitHub

Stop making models relearn the repository.

Give them the project truth they actually need.

Atlas compiles current repository evidence into a bounded working set any coding agent can use.

Inspect the system Open GitHub
Layered paper, glass, and metal arranged into a compact evidence stack
Evidence plate 01 Repository truth compressed into a task-bounded view.

01Local-firstProject source stays under your control.

02Provider-neutralOne evidence layer for any model or agent.

03CLI + MCPShared contracts across human and agent workflows.

04Dual licensedMIT or Apache-2.0.

01THE OPERATING PROBLEM

Every new session pays the rediscovery tax.

Agents list directories, search broad terms, reopen the same files, and reconstruct relationships that the last session already found.

The cost is larger than tokens.

Partial scans and stale assumptions become confident changes.

Atlas makes project understanding persistent, qualified, and bounded to the task at hand.

02SYSTEM ARCHITECTURE

Compile project truth once. Serve only what the task needs.

Atlas keeps verified evidence persistent, then assembles a small working set from one immutable generation.

  1. 01 / INPUT Repository state
    • Source
    • History
    • Tests
    • Diagnostics
  2. 02 / QUALIFY Truth Plane

    Files, revisions, symbols, relationships, effects, coverage, conflicts, and lifecycle history.

  3. 03 / COMPILE Context IR

    Generation-bound evidence with selection reasons, costs, omissions, and uncertainty intact.

  4. 04 / SERVE Any model or agent

    Works from current evidence without rescanning the whole repository.

Minimum sufficient working setQualified project evidence in. Bounded context out.

03EVIDENCE DISCIPLINE
Technical papers sorted from broad research into a precise central evidence set
Evidence plate 02 / Broad repository signals resolved into a precise working set.

Evidence stays qualified.

Atlas does not replace source, builds, tests, or runtime evidence. It records where evidence came from and whether it is current, verified, unresolved, conflicting, stale, omitted, or unavailable.

CurrentVerifiedUnresolvedConflictingStaleOmittedUnavailable
Freshness
Live source checks block stale change-mode use.
Boundaries
Fixed budgets keep task context controlled and inspectable.
Failure
Failed candidates never partially replace the last valid generation.
04ENGAGEMENT PATHS

One evidence contract. Three clear ways in.

Start in source, standardize context across an engineering organization, or test the research question in a real workflow.

01 / Open source users

Build locally. Inspect every contract.

The release candidate is source-built, local-first, provider-neutral, and dual licensed. Public repository access is being prepared.

Open GitHub (access required)
LOCAL / SHELL
cargo build --release --locked
atlas init <repository>
atlas reconcile <repository>

02 / Engineering teams

Integrate one evidence contract across tools.

Use CLI and MCP surfaces to share deterministic, generation-bound context without coupling Atlas to one model provider.

Review operations

Deterministic context, shared across human and agent workflows.

03 / Commercial and research partners

Test the context question that matters.

Run pilots around context yield, accepted-task usefulness, and integration with real engineering workflows. No model lock-in is required.

Contact project owner

Pilot the system against real repositories and real engineering constraints.

05RELEASE BOUNDARY

Shipped capabilities stay separate from the research horizon.

The V1.1 to V1.4 current candidate provides persistent project truth, deterministic task compilation, temporal intelligence, and CLI/MCP parity. V1.5 and V2.0 remain future work.

VersionCapabilityStatus
  • V1.1Semantic TruthCurrent candidate
  • V1.2Context IR + ObservabilityCurrent candidate
  • V1.3Deterministic Task CompilerCurrent candidate
  • V1.4Temporal IntelligenceCurrent candidate
  • V1.5Working-Set OptimizationFuture work
  • V2.0Project Context CompilerFuture work
06RESEARCH QUESTION

How small can a model's working view become without reducing correctness?

Workspace Atlas turns that question into an inspectable engineering system.

Open GitHub Partner with Atlas