Skip to main content
Glama
Architect Agent Framework MCP Server

Architect Agent Framework MCP Server

Architect Agent Framework plugin

Candidate: 1.0.0. Release status: hihi.

English · Tiếng Việt · Install Guide · Hướng dẫn cài đặt · Contributing · Hướng dẫn đóng góp · Changelog

Use the plugin to prepare bounded architecture tasks in a current core workspace across Claude Code (claude), OpenAI Codex (codex), and DeepSeek Harness (dsh). It packages four catalog skills, three foundation skills, and one runtime skill, alongside an embedded zero-install runtime engine (architect_framework/, vendor/sdd-agent/, and contracts/). Its local MCP server exposes proposals, scheduling, and observation.

Use the target host's local plugin procedure. Configure the launcher with this checkout's interpreter. The alpha package is not a published release.

The MCP and hook manifests require a Python 3.10+ command named python on the host's PATH, including native Windows. Set ARCHITECT_PYTHON to the prepared environment's interpreter. Otherwise the launcher probes its current interpreter and uses uv, when available, to provide missing dependencies in an isolated environment. Run any runtime action directly through python scripts/architect_runtime.py --bootstrap <command> from the plugin directory.

A missing marker requires explicit setup or migration. A malformed marker requires repair. Old approvals do not authorize current work.

Quickstart and multi-harness installation

Install the bundled skills, shared templates, and 16 commands across Claude Code, OpenAI Codex, and DeepSeek Harness directly from this repository:

./install.sh --all --agent all        # macOS / Linux / WSL / Git Bash
.\install.ps1 -All -Agent all        # Windows PowerShell

Copy .env.example to .env (or ~/.architect-framework/.env) when configuring an optional live judge API key (JEV_API_KEY). See INSTALL.md and INSTALL.vi.md for full setup details.

Related MCP server: NS Workflow

Eight end-to-end architecture workflows (UC-01 .. UC-08)

Use Case

Command

Default Profile

Outcome

UC-01: Raw Idea to Clarified SRS

/architect-clarify <idea or brief.md>

seed · solo-author

Socratic interview (max 5 questions with defaults) producing clarified-requirements.md

UC-02: Fast-Track MVP Specification

/architect-quickstart [source]

standard · solo-author

Materializes 12–25 core pages, stamps witnesses, and passes sdd check

UC-03: Full Enterprise Maximal Spec

/architect-design --profile maximal --sources <...> --out <dir>

maximal · solo-author

Autonomous 6-phase execution (Intake, Materialize, Waves, Prepare-Check, Gate, Review)

UC-04: Brownfield Code & Test Audit

/architect-audit <code/test> --against <SRS>

standard / maximal

Reconciles implementation against specs and records CNF-* conflicts

UC-05: Point Solution (Single ADR/API)

/architect-route & /architect-execute --standalone

standalone

Routes to Domain × Task skill pair and writes a standalone artifact

UC-06: Spec Critique & Gate Check

/architect-check --full & /architect-review [--live]

Current workspace

Runs anti-stub/depth gates (DTG-01..10), 02-spec-critic, and sharded judge

UC-07: Living Spec Change Request

/architect-change <change description>

Incremental

Computes graph impact radius and verifies DTG-07 non-regression

UC-08: C4 Console, Skills & Sign-Off

/architect-console, /architect-skills, /architect-approve

Workbench (:8767)

Interactive C4 map (< 20ms reader), 297 skill slot manager, and WebAuthn Passkey sign-off

Read the complete 16-command manual and multi-harness guide in docs/QUICKSTART-USE-CASES.md.

Workbench and human decisions

Open the printed localhost link. /runs shows Overview, Runs, Gate inbox, Documents, and Settings. / shows exact candidate changes and document decisions.

A dispatch signature permits one task attempt. Document acceptance remains separate. Reading starts no work. A scheduling pause stops new dispatch. Running host workers can continue. Cancellation fences late results.

The person starts architect-runtime console serve <workspace> --enroll in their own terminal. Register a configured reviewer and confirm the fingerprint there. Copied keys need explicit trust on each machine.

The agent cannot enroll keys, sign decisions, or accept documents. MCP exposes no human approval tool. Required gates stay pending without a trusted key.

Review and installed skills

Configure an exact model, scope, deadline, attempts, and reservation budget. Inspection does not invoke it. Screening cannot certify grounded correctness. A label signature does not qualify a model automatically.

Use architect-runtime skills offer <skill-id> to inspect a slot. Use architect-runtime skills install <package.zip> to install compatible content. Accepted grants control writes.

Read the bundled runtime skill and command catalog. The writing checker is partial and does not certify ASD-STE100.

Current core paths replace the retired runtime. Release qualification remains incomplete where independent human evidence is absent. Local tests use synthetic workers and virtual keys. Live hosts, physical devices, native OS runs, real models, and user studies require separate evidence.

Current workbench preview

These pictures use the public payments example. They show pending synthetic tasks, not accepted documents. These previews are synthetic workbench examples. The evidence index records the verified source and artifact revision.

Open the local workbench overview to inspect pending task counts and incomplete readiness.

Use /runs for dependencies, write scope, and partial observation.

Review task scope in the dispatch gate. Document acceptance remains a separate decision.

Packaged surfaces

Surface

Count

Commands

16

Skills

8

MCP tools

29

MCP tools expose core planning, candidates and observation. Human signatures stay in the trusted review surface.

Documentation, issue reporting, and community contributions

Available Tools

29 tools
architect_candidate_reviewA

Review an exact candidate: checks, semantic judge and all-document language evidence. live=true explicitly permits configured paid calls within the shared budget. Never accepts.

Inputs: change_id (required), workspace (required), live, force.

architect_capabilitiesB
Read-onlyIdempotent

Read supported host actions, coverage and authority boundaries.

Inputs: no named parameters.

architect_console_urlA
Read-onlyIdempotent

The current workbench link for task gates and document acceptance. Hand it to the user; they read and sign there with a passkey. Read-only: no tool decides.

Inputs: workspace (required), port, wait.

architect_doctorC
Read-onlyIdempotent

Read bounded redacted diagnostics without modifying run state.

Inputs: workspace (required).

architect_execute_standaloneC

Execute a single Domain x Task skill pair in standalone mode (UC-05) without requiring sdd-workspace.yml.

Inputs: query (required), out, task, domain, +4 more.

architect_language_repairB

Propose a targeted host-authored repair under an accepted grant. Preserves facts and live files; at most two repair attempts per document lineage. Candidate still needs review.

Inputs: text (required), skill (required), change_id (required), workspace (required), +3 more.

architect_language_reviewA

Record revision-bound per-rule language findings. live=true sends complete context to configured Jev; false runs local checks. No qualification or approval authority.

Inputs: change_id (required), workspace (required), live, force.

architect_runs_controlC

Pause, resume, cancel or retry scheduling; no approval authority.

Inputs: action (required), run_id (required), workspace (required), task_id.

architect_runs_observeA
Read-onlyIdempotent

Read durable core events and task snapshots; starts no work.

Inputs: workspace (required), after, database_id.

architect_runs_startC

Start a bounded, revision-bound core task plan; human decisions remain separate.

Inputs: tasks (required), workspace (required), run_id, base_change.

architect_runtime_nextC

What to do now: jobs to execute (with brief and allowed paths) and what waits for a human.

Inputs: workspace (required), run.

architect_runtime_planB
Read-onlyIdempotent

Preview a revision-bound task plan and gates without creating a run.

Inputs: workspace (required), only, tasks, base_change.

architect_runtime_recheckB

Re-read the ledgers after a human decided. Carries no authority: an unapproved wait just waits again.

Inputs: workspace (required), run.

architect_runtime_scorecardC
Read-onlyIdempotent

Read the current core task summary; document acceptance is separate.

Inputs: workspace (required), run.

architect_runtime_skillsA
Read-onlyIdempotent

The skill slots this workspace needs: which have content, which are empty, what fills the empty ones (pack, price, store) and how to install a bought zip. Same JSON as skills status --json.

Inputs: workspace (required).

architect_runtime_startB

Start a run in harness mode: the host agent is the executor. Returns what waits.

Inputs: workspace (required), only.

architect_runtime_statusC
Read-onlyIdempotent

Run state, per-agent lifecycle state and every pending wait.

Inputs: workspace (required), run.

architect_runtime_tickC

Continue notify waits whose deadline passed.

Inputs: workspace (required), run.

architect_sdd_checkC
Read-onlyIdempotent

Read deterministic checks, regression readiness and freshness without accepting or changing state.

Inputs: workspace (required), change_id.

architect_sdd_extractB

Extract source requirements, ADRs, scenarios, and conflict records into .architect-build/harvest.json (auto-discovers sources when omitted).

Inputs: workspace (required), audit, sources.

architect_sdd_initB

Initialize a new SDD workspace in the target directory (supports solo-author and governed-team presets and existing source files).

Inputs: workspace (required), preset, profile, reviewer, +2 more.

architect_sdd_materializeC

Materialize the target architecture document tree, ledgers, wave micro-batches, and frozen basis_digest.

Inputs: workspace (required), preset, harvest, profile, +4 more.

architect_sdd_prepare_checkB

Run the built-in post-processing pipeline (strip , tighten lines, repair frontmatter, STE-80 style fix, stamp depth_witnesses, freeze digest).

Inputs: workspace (required), stamp, tighten, style_fix, +1 more.

architect_sdd_proposeC

Create an isolated candidate under a configured skill grant and expected live snapshot. Never accepts.

Inputs: skill (required), workspace (required), operations (required), expected_snapshot (required), +2 more.

architect_sdd_readC
Read-onlyIdempotent

Read a bounded registered artifact slice with its exact live or candidate revision.

Inputs: workspace (required), artifact_id (required), limit, offset, +1 more.

architect_style_checkB
Read-onlyIdempotent

Check declared STE-inspired subset for one exact registered revision; incomplete rules stay unknown. No approval authority.

Inputs: workspace (required), artifact_id (required), change_id.

architect_task_claimC

Claim a ready core task and its private attempt lease.

Inputs: run_id (required), task_id (required), workspace (required).

architect_task_completeC

Submit scoped proposal operations; never accepts or publishes documents.

Inputs: run_id (required), task_id (required), workspace (required), operations (required), +1 more.

architect_task_heartbeatC

Renew the exact active attempt lease.

Inputs: run_id (required), task_id (required), workspace (required), lease_token (required).

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    C2.8/5.0

    Scored across 29 tools

    Disambiguation3/5

    The tool set has many similarly named lifecycle phases, especially architect_runs_start vs architect_runtime_start/plan, architect_sdd_check vs architect_style_check vs architect_candidate_review, and architect_runtime_status vs architect_runs_observe. Descriptions do clarify most boundaries, but an agent could still misselect among the runtime/run/review/check tools without careful reading.

    Naming Consistency4/5

    Most tools follow an architect_<domain>_<action> snake_case pattern, with clear groupings for sdd, runtime, runs, task, and language. There are minor deviations such as architect_capabilities, architect_doctor, architect_console_url, and noun-first review names, but the overall convention is predictable.

    Tool Count2/5

    With 29 tools, the surface is heavy and exceeds the typical well-scoped range. The domain is complex, but several tools cover narrow runtime phases or review/check states, making the overall set cumbersome for an agent to navigate.

    Completeness4/5

    The surface covers the SDD lifecycle from init and extraction through materialization, proposal, review, runtime execution, task leasing, diagnostics, and human handoff via the console URL. Minor gaps exist around explicit installation or broader workspace administration, but core workflows and human acceptance boundaries are represented.

    Maintenance

    ActivityNo data
    ResponsivenessSyncing

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      C
      quality
      B
      maintenance
      Enables AI agents to collaborate via shared blackboard state, multi-agent deliberation, human-in-the-loop proposal approval, operating-mode control, and audit logging through MCP.
      11
      -
    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables MCP-capable AI agents to work on bounded project folders through read, write, run, and verify tools, with durable approval/resume workflows and completion proven by observed verification evidence rather than self-reported results.
      MIT
    • A
      license
      Not graded
      quality
      C
      maintenance
      Provides MCP tools for AI agents to create and read decision decks, propose cards and build plans, and retrieve finished tasks, enabling local decision review workflows with portable task handoffs.
      2
      MIT
    • A
      license
      Not graded
      quality
      A
      maintenance
      Enables structured multi-agent workflows across AI coding providers by managing documents, tracked tasks, handoffs, health evidence, and MCP-owned task/action state. It supports spec creation, implementation coordination, review, triage, and health-gated completion.
      29 npm
      MIT