Sector Radar
VS Code AI Coding Agent Extensions
Published by Diversum. Denominator: 22 in-cluster accepted candidates. Evidence snapshot: 2026-05-17.
Executive Summary
The VS Code AI Coding Agent Extensions sector has a recognizable default: a TypeScript or JavaScript VS Code-compatible extension with a sidebar, chat, or webview surface, package-backed distribution, recent activity, and AI coding-agent behavior that reaches beyond static chat.
The sector is top-heavy. Cline, Continue, and Kilo define much of the current attractor by adoption and engineering depth, while the long tail explores Claude Code bridges, ACP clients, MCP debugging, local models, markdown chat, prompt/apply workflows, and multi-agent orchestration.
The most important structural finding is the bridge pattern. The sector shares MCP/session/local-process traits with MCP Memory Servers and shares TypeScript, webview UI, extension packaging, marketplace distribution, and fork-and-build lineage with Chrome Extension Starters.
Three retained candidates are Adjacent Infrastructure inside the corpus: cogflows/promptcode-vscode, jasonjmcghee/claude-debugs-for-you, and formulahendry/vscode-acp. They remain visible because they show where prompt tooling, MCP clients, and ACP clients touch the agent-extension boundary without becoming the default agent runtime.
About This Report
Who We Are
Diversum maps convergence patterns in software ecosystems. A Sector Radar identifies the default project shape in a sector, measures how public repositories relate to that default, and surfaces where meaningful differentiation still exists.
This report applies the same public Sector Radar structure used for MCP Memory Servers and Chrome Extension Starters. Presentation summarizes accepted SC-024 artifacts; it does not change the corpus, scores, centroid traits, or methodology.
Conflict of Interest Disclosure
Diversum has no affiliation with any repository in this corpus. No project in this report is built or maintained by the report authors.
Evidence Snapshot
All scores, roles, source observations, star counts, and freshness facts are frozen at the 2026-05-17 evidence snapshot. Repositories can change after publication.
Correction Invitation
Correction route: send a note to hello@diversum.dev with the repo name, displayed value or claim, report section, and source evidence that should replace it.
Methodology
Key Terms
- Synthetic centroid - The composite default project assembled from traits shared by most in-cluster repositories. It is not a real project and not a quality ideal.
- In-cluster - Public, fresh, non-archived, package-backed VS Code or VS Code-compatible extension repositories whose headline behavior is AI coding assistant, coding agent, or coding-agent client workflow.
- Adjacent Infrastructure - Retained boundary-visible projects whose architecture matters to the sector but does not define the default IDE-agent extension pattern.
Denominator Protocol
The final Denominator is In-cluster, N=22. SC-021 source verification accepted 22 candidates, and SC-024 refreshed GitHub metadata plus local repo evidence packets before scoring. Issue #19 remains scout context only; metadata-only candidates from that thread are not promoted into this report.
Five Radar Dimensions
| Dimension | What it measures | Reader meaning |
|---|---|---|
| Centroid Match (CM) | How closely a repo matches the 10-trait sector centroid. | Whether the repo looks like the sector default. |
| Differentiation (DF) | Observable traits or combinations beyond the centroid. | Whether the repo has a distinct position. |
| Human-Edge (HE) | Sustained engineering depth visible in code, package structure, docs, tests, and maintenance. | Whether there is reviewable human engineering depth. |
| Adoption Depth (AD) | Stars, forks, issue/use signals, and ecosystem visibility. | Whether public reach is strong or thin. |
| Claim/Implementation Integrity (CI) | Whether README and manifest claims trace to package and code evidence. | Whether public claims match source evidence. |
Public Cluster Roles
| Role | Rule basis | What it means |
|---|---|---|
| Category Leader | HE >= 4, AD >= 4, CM >= 3 | High engineering depth, adoption, and centroid alignment. |
| Differentiated Niche | DF >= 3 and HE >= 2 | In-cluster, but meaningfully distinct from the default. |
| Strong Default | CM >= 4, DF <= 2, HE >= 2 | Recognizable default implementation with limited independent differentiation. |
| Baseline Implementation | CM >= 3 and DF <= 1 | Active package-backed implementation with thin independent signal. |
| Adjacent Infrastructure | CM <= 2 and HE >= 3 | Boundary-visible infrastructure relevant to the sector edge. |
| Inactive Scaffold | HE <= 1 | Structurally present but low sustained engineering signal. No retained repo has this role in this corpus. |
The Sector
What VS Code AI Coding Agent Extensions Are
This sector covers public GitHub repositories with source-visible VS Code or VS Code-compatible extension package evidence where the headline product is an AI coding assistant, coding agent, coding-agent client, or IDE-native coding workflow surface.
The boundary excludes closed-source commercial-only agents, whole-editor forks, generic VS Code samples, standalone CLIs or web services with no VS Code package surface, and protocol/runtime tools unless they also ship the coding-agent or coding-agent client extension surface.
Why This Sector Exists
AI coding agents increasingly meet developers inside the editor. VS Code extension packaging, webview or sidebar surfaces, local tool bridges, configurable providers, and workspace operations create a sector shape that is different from standalone CLIs, generic assistant extensions, MCP servers, or browser-extension starters.
The Synthetic Centroid
The default project is a TypeScript or JavaScript VS Code extension that presents an AI coding assistant through a sidebar, chat, or webview surface. It is package-backed, fresh and non-archived, individual- or small-team-led, and framed around file/workspace context, edits, tools, local or external runtime bridges, and configurable model providers.
Source: Manual-audit/CENTROID-VS-CODE-AI-CODING-AGENT-EXTENSIONS.md. Denominator: In-cluster, N=22. Reader meaning: trait prevalence shows what defines the sector default, not what all projects must become.
| Trait | Name | Prevalence | Reader meaning |
|---|---|---|---|
| C1 | VS Code package surface | 22/22 (100%) | Every retained repo has source-visible extension package evidence. |
| C2 | TypeScript / JavaScript extension implementation | 19/22 (86%) | The default implementation stack is TS/JS/Node. |
| C3 | Sidebar, chat, or webview IDE surface | 17/22 (77%) | The common user surface is IDE-native rather than external-only. |
| C4 | AI coding assistant or coding-agent headline | 22/22 (100%) | Coding-agent relevance is an inclusion-gate trait. |
| C5 | Individual or small-team maintenance | 16/22 (73%) | The sector is mostly independent or small-team maintained. |
| C6 | Multi-model or BYO provider posture | 15/22 (68%) | Provider flexibility is common but not universal. |
| C7 | Agentic file/workspace operation | 16/22 (73%) | The default goes beyond static chat into workspace actions. |
| C8 | Local tool/process bridge | 14/22 (64%) | MCP, ACP, CLI, local-model, debugger, or process bridges shape the edge. |
| C9 | Recent active source | 22/22 (100%) | Freshness is preserved by the accepted corpus boundary. |
| C10 | Marketplace/installable extension package | 22/22 (100%) | The sector is package-backed, not documentation-only. |
Cluster Map
Source: Manual-audit/CALIBRATION-TABLE-RADAR-VS-CODE-AI-CODING-AGENT-EXTENSIONS.md and Reports/SECTOR-RADAR-VS-CODE-AI-CODING-AGENT-EXTENSIONS.md. Denominator: In-cluster, N=22. Reader meaning: roles describe sector position at the evidence snapshot, not maintainer worth or universal product quality.
Category Leaders (5)
cline/cline, continuedev/continue, Kilo-Org/kilocode, zgsm-ai/costrict, and codota/tabnine-vscode combine high adoption, high engineering depth, and strong centroid alignment. Cline and Continue define the practical category shape; Kilo adds lineage/fork-swarm evidence; costrict shows enterprise-oriented agent packaging; Tabnine keeps mature completion/chat-extension lineage visible through its extension-specific repo.
Differentiated Niches (7)
DeepMyst/Mysti, andrepimenta/claude-code-chat, adifyr/shadow-code, rusiaaman/chat.md, Hive-Academy/ptah-extension, AizenvoltPrime/damocles, and joaompfp/hermes-vscode are structurally distinct through multi-agent orchestration, Claude Code bridges, markdown-as-interface, pseudocode-first workflows, Agent SDK/subagent surfaces, and ACP-based agent clients.
Strong Defaults (4)
Lianues/Lim-Code, vscode-reborn-ai/vscode-reborn-ai, Kile-Thomson/Rokket-GSD, and Harsh1210/openclaude-vscode implement recognizable default patterns without enough unique architecture to define a separate niche.
Baseline Implementations (3)
10Nates/ollama-autocoder, ZhouChaunge/DeepCopilot, and justimyhxu/claude-code-local-for-vscode are active package-backed implementations with limited independent signal at the snapshot date.
Adjacent Infrastructure (3)
cogflows/promptcode-vscode, jasonjmcghee/claude-debugs-for-you, and formulahendry/vscode-acp are source-backed and relevant to the sector boundary, but their architecture is not the default agent-extension pattern.
Interactive Evidence Graph
This graph turns the SC-024 centroid, calibration table, visual data, and evidence refresh into an explorable evidence surface. Switch modes to see repositories connected by shared centroid traits, protocol/runtime surfaces, workflow themes, lineage/client-family notes, or adjacent boundary signals.
Graph proximity and edges do not imply causation, authorship, derivation, dependency, or endorsement. Layout coordinates are presentation choices; edge reasons and evidence references are the traceable fields.
Sector Health
Source: Reports/VISUAL-DATA-VS-CODE-AI-CODING-AGENT-EXTENSIONS.md Table 4 and the calibration score matrix. Denominator: In-cluster, N=22 unless the row names a count. Reader meaning: health metrics summarize convergence, differentiation, engineering depth, adoption, integrity, and boundary pressure.
| Metric | Value | Scale | Denominator / sample | Reader meaning |
|---|---|---|---|---|
| Convergence Index | 3.55 | 0-5 | In-cluster, N=22 | Moderate-high convergence around package-backed IDE agent surfaces. |
| Differentiation Index | 3.05 | 0-5 | In-cluster, N=22 | Protocols, runtimes, and workflow shapes keep the sector meaningfully varied. |
| Human-Edge Index | 3.09 | 0-5 | In-cluster, N=22 | Strong anchors coexist with a thinner long tail. |
| Adoption Index | 2.36 | 0-5 | In-cluster, N=22 | Adoption is concentrated in a small number of anchor repos. |
| Integrity Index | 3.64 | 0-5 | In-cluster, N=22 | Claims generally trace to source package evidence. |
| Viability Rate | 72.7% (16/22) | % | 16 of 22 retained repos | Sixteen repos show HE >= 3. |
| Inactive Scaffold Rate | 0.0% (0/22) | % | 0 of 22 retained repos | No retained repo is classified as Inactive Scaffold. |
| Category Leakage Rate | 13.6% (3/22) | % | 3 of 22 retained repos | Three retained boundary cases are classified as Adjacent Infrastructure. |
Key Findings
1. The sector has a strong agent-extension attractor
The default is not a generic chat extension. The active attractor is an IDE-native agent surface with file/workspace context, tool or process bridges, configurable providers, and package-backed distribution.
2. Adoption is extremely top-heavy
Cline, Continue, and Kilo account for the overwhelming majority of stars and forks. The rest of the corpus remains strategically useful, but most long-tail repos have AD scores of 1-2.
3. MCP and ACP are boundary-shaping, not universal
MCP appears through Cline-style tooling, markdown chat, debugging bridges, permissions packages, and orchestration surfaces. ACP appears through formulahendry/vscode-acp and joaompfp/hermes-vscode. The report keeps them separate because the architectures imply different client/runtime relationships.
4. Lineage is a sector trait, not a hidden override
Kilo and justimyhxu/claude-code-local-for-vscode carry lineage-bearing notes. Both remain active package-backed repos in the source refresh, so they are retained. The lineage note affects interpretation and role assignment; it is not a hidden shortcut.
5. The sector contributes future bridge evidence
This report does not publish a Super Cluster graph. It records why the sector can contribute later evidence: MCP/session/local-process traits connect in one direction, while TypeScript/webview/extension-packaging traits connect in another.
Differentiation Gaps
| Gap | Why it matters | Source basis |
|---|---|---|
| Trustable local-first agent memory inside VS Code | MCP and local-model pieces exist, but durable memory/session persistence is rarely the core IDE-agent product. | Report differentiation gaps; centroid C8 evidence. |
| Protocol-neutral agent client UX | ACP and MCP client surfaces are emerging, but mature multi-protocol client UX is not the common pattern. | Protocol/runtime distribution and boundary notes. |
| Auditable claim surfaces | The long tail often claims broad agent capability without enough tests, examples, or issue history to make claims easy to verify. | CI scores and calibration evidence notes. |
| Open evaluation loops | Evaluation direction appears in anchors, but reproducible public evals are not a corpus-wide norm. | Report gaps and row-level evidence notes. |
Sector Narrative
The VS Code AI Coding Agent Extensions sector is past the generic assistant phase. The sector default now expects installable extension packaging, a visible IDE surface, source evidence of coding-agent positioning, and at least some connection to workspace actions, local tools, external runtimes, or provider configuration.
That convergence does not erase variation. The strongest differentiation appears at the runtime and workflow edges: local models, MCP debugging, ACP clients, Claude Code bridges, markdown-first interaction, pseudocode workflows, and multi-agent orchestration. Those edges explain why the sector is more differentiated than Chrome Extension Starters even though the extension packaging shape is stable.
The public interpretation should therefore avoid a single-winner story. The useful pattern is an attractor with boundary pressure: anchor projects define the practical default, while adjacent and niche projects show where the next interface, protocol, or workflow differences may matter.
Appendix: Scoring and Source Summary
Source Artifacts
Source map: this page summarizes accepted SC-024 artifacts without changing their analysis. Found a stale repo, wrong source trace, or boundary classification issue? Send a correction to hello@diversum.dev with the repo name, displayed value, and replacement source.
| Published artifact | Source role | Visible metric families |
|---|---|---|
| Reports/SECTOR-RADAR-VS-CODE-AI-CODING-AGENT-EXTENSIONS.md | Public narrative report | Executive summary, findings, gaps, narrative, limitations. |
| Reports/VISUAL-DATA-VS-CODE-AI-CODING-AGENT-EXTENSIONS.md | Chart-ready aggregations | Role composition, health metrics, protocol surfaces, score matrix. |
| Manual-audit/CENTROID-VS-CODE-AI-CODING-AGENT-EXTENSIONS.md | Centroid traits and corpus boundary | Centroid definition, trait prevalence, sensitivity notes. |
| Manual-audit/CALIBRATION-TABLE-RADAR-VS-CODE-AI-CODING-AGENT-EXTENSIONS.md | Scores, roles, rule traces, evidence notes | Cluster map, health metrics, role assignment, boundary register. |
| Operations/Sprint-Contracts/SC-024.md | Original report-generation sprint scope | Public report acceptance context. |
| Operations/Sprint-Contracts/SC-025.md | Format-parity sprint scope | Report page parity and validation requirements. |
All 22 Repositories
| Repo | CM | DF | HE | AD | CI | Role |
|---|---|---|---|---|---|---|
| cline/cline | 5 | 5 | 5 | 5 | 5 | Category Leader |
| continuedev/continue | 5 | 5 | 5 | 5 | 5 | Category Leader |
| Kilo-Org/kilocode | 5 | 4 | 5 | 5 | 4 | Category Leader |
| zgsm-ai/costrict | 4 | 4 | 4 | 4 | 4 | Category Leader |
| codota/tabnine-vscode | 3 | 3 | 4 | 4 | 4 | Category Leader |
| DeepMyst/Mysti | 4 | 4 | 3 | 3 | 4 | Differentiated Niche |
| 10Nates/ollama-autocoder | 3 | 1 | 2 | 2 | 3 | Baseline Implementation |
| Lianues/Lim-Code | 4 | 2 | 3 | 2 | 3 | Strong Default |
| andrepimenta/claude-code-chat | 4 | 3 | 3 | 4 | 4 | Differentiated Niche |
| adifyr/shadow-code | 3 | 4 | 3 | 2 | 3 | Differentiated Niche |
| rusiaaman/chat.md | 3 | 4 | 3 | 2 | 3 | Differentiated Niche |
| vscode-reborn-ai/vscode-reborn-ai | 4 | 2 | 3 | 2 | 3 | Strong Default |
| cogflows/promptcode-vscode | 2 | 3 | 3 | 1 | 3 | Adjacent Infrastructure |
| ZhouChaunge/DeepCopilot | 3 | 1 | 2 | 1 | 3 | Baseline Implementation |
| Kile-Thomson/Rokket-GSD | 4 | 2 | 2 | 1 | 3 | Strong Default |
| jasonjmcghee/claude-debugs-for-you | 2 | 3 | 3 | 2 | 3 | Adjacent Infrastructure |
| Hive-Academy/ptah-extension | 4 | 4 | 3 | 1 | 4 | Differentiated Niche |
| AizenvoltPrime/damocles | 4 | 3 | 3 | 1 | 4 | Differentiated Niche |
| justimyhxu/claude-code-local-for-vscode | 3 | 1 | 2 | 1 | 3 | Baseline Implementation |
| formulahendry/vscode-acp | 2 | 4 | 3 | 2 | 4 | Adjacent Infrastructure |
| joaompfp/hermes-vscode | 3 | 3 | 2 | 1 | 4 | Differentiated Niche |
| Harsh1210/openclaude-vscode | 4 | 2 | 2 | 1 | 4 | Strong Default |
Methodology Notes
Methodology: Sector Radar v1 with source-verified intake, accepted N=22 corpus, and public-safe cluster labels. Canonical scoring definitions remain in Manual-audit/AUDIT-TEMPLATE-SECTOR-RADAR-v1.md.