Skip to content

Best MCP Servers for GitHub and Engineering in 2026

Official GitHub MCP plus the servers you wire next to it for merge requests, issues, errors, CI, and review.

As featured inTechCrunchForbesBloombergBusiness InsiderThe Verge
868 AI Agents tools tracked
TL;DR

GitHub MCP Server is the first install: issues, pull requests, repos, and Actions from the agent you already use. Wire GitLab MCP Server if that is the host. Put Sentry MCP on the error path, Linear MCP or Jira MCP Server on the ticket path, Semgrep MCP on the diff, and CircleCI MCP on the pipeline. Context7 MCP feeds current library docs so the agent stops inventing APIs. Figma MCP is handoff, not git. Exa MCP is code and web search when the answer is not in the repo. Toolradar tracks 373 published tools that are MCP servers or MCP-named in the name, slug, tagline, or description. The AI Agents category holds 856 published tools, 56% of them free or freemium.

The twelve thematic MCP guides already on Toolradar cover databases, analytics, devops, search, comms, finance, sales, productivity, project management, ecommerce, support, and file storage. None of them is the git, pull request, and issue loop. That is this list.

GitHub MCP Server is pick one because that is where most engineering work already lives: the PR, the review comment, the failing check, the issue that should have been linked. An agent that can read a repo but cannot open a pull request is a chatbot with extra steps. The servers after it are the ones you attach to that loop. GitLab MCP Server if the org never moved off GitLab. Sentry MCP when the question is "what broke in production." Linear MCP or Jira MCP Server when the question is "what is the ticket." Semgrep MCP when the question is "did this diff introduce a finding." CircleCI MCP when the question is "why is main red." Context7 MCP when the question is "what does this library actually export in 2026." Figma MCP when the question is "what did design spec," which is handoff, not version control. Exa MCP when the answer is on the web or in someone else's public code.

Toolradar tracks 373 published tools that are MCP servers or MCP-named (name, slug, tagline, or description). The AI Agents category has 856 published tools, and 56% of that category is free or freemium. This guide does not invent a larger catalog number. It ranks ten servers by how they sit next to GitHub in a normal engineering week.

Top Picks

Based on features, user feedback, and value for money.

ToolStarting priceRatingBest for
GitHub MCP ServerFreen/aAny team whose code review and issues already live on GitHub.
GitLab MCP ServerFreen/aOrgs on GitLab.com or self-managed GitLab that want the same agent loop GitHub shops get.
Sentry MCPFreen/aTeams that triage in Sentry and want the agent to read the event, not a pasted stack tr...
Linear MCPFree4.5(89)Product-engineering teams whose source of truth for work items is Linear.
Context7 MCPFrom $10/mon/aAny repo that depends on frameworks whose APIs moved after the model's training cutoff.
Jira MCP ServerFree4.4(23,083)Enterprise and Jira-native teams who want the agent on the ticket, not on a shadow GitH...
Semgrep MCPFree4.6(55)Teams that already run Semgrep in CI and want those findings inside the agent session.
CircleCI MCPFrom $15/mon/aRepos whose required checks run on CircleCI rather than GitHub Actions alone.
Figma MCPFrom $3/mo4.6(2,863)UI work where the agent should see the file the designer named, not a screenshot in the...
Exa MCPFrom $1/mon/aAgents that need a current example, a docs page, or a public snippet the repo does not...

Any team whose code review and issues already live on GitHub.

+Official server, so the tool names track GitHub's own API instead of a hobby wrapper
+Covers the objects you already use: issues, PRs, comments, checks, and repo contents
+Works with a fine-grained token so you can limit it to the repos the agent should see
It does not see Sentry, Linear, or CircleCI. Those are other servers on this list
A classic PAT with repo scope is too wide; spend the time on fine-grained permissions
Great value

The GitHub MCP Server pricing is exceptionally generous, offering a full-featured 'Open Source' tier completely free.

Watch out

Requires self-hosting infrastructure

Orgs on GitLab.com or self-managed GitLab that want the same agent loop GitHub shops get.

+Maps onto merge requests and issues, not a GitHub-shaped API with different nouns
+Pipeline status belongs next to the MR the same way Actions belong next to a PR
+Self-managed GitLab is in play, which matters for companies that will not put code on github.com
Do not run **GitHub MCP Server** and **GitLab MCP Server** pointed at the same work. Pick the host you use
Permission models differ from GitHub; copy-pasting a token policy will over-scope it

Teams that triage in Sentry and want the agent to read the event, not a pasted stack trace.

+Turns a production blow-up into a tool call instead of a tab hunt
+Issue context next to **GitHub MCP Server** is how a fix PR cites a real event ID
+Traces and breadcrumbs beat a single-line error message in a chat log
Noisy projects will flood the agent. Filter by environment and project first
It will not deploy the fix. That is still CI plus a human
Great value

The pricing for Sentry MCP is exceptionally generous, offering a full-featured 'Open Source' tier completely free.

Watch out

Requires self-hosting infrastructure

4
Linear MCP logo

Linear MCP

4.5G2(86)4.3Capterra(3)

Product-engineering teams whose source of truth for work items is Linear.

Linear MCP screenshot
+The agent can attach a PR to the issue that actually got scheduled this cycle
+Cycles and project state stop the model from inventing priority
+Faster than asking someone to paste a Linear URL into the prompt every time
If the org lives in Jira, install **Jira MCP Server** instead of both
An agent that can create issues will create noise unless you constrain the project
5
Context7 MCP logo

Context7 MCP

5.0G2(1)

Any repo that depends on frameworks whose APIs moved after the model's training cutoff.

Context7 MCP screenshot
+Grounds answers in current docs, which is the failure mode of asking for a Prisma query from memory in 2026
+Works beside **GitHub MCP Server**: the agent reads your code and the library, not one or the other
+Cuts a class of review comments that exist only because the model remembered last year's API
It is not a git host. It will not open the PR
You still have to name the library and version you care about
Good value

Context7 MCP offers a fair pricing structure, with a generous free tier for individual developers.

Watch out

Pro tier overage: $10 per 1,000 extra API calls.

6
Jira MCP Server logo

Jira MCP Server

4.4Capterra(15,362)4.3G2(7,721)

Enterprise and Jira-native teams who want the agent on the ticket, not on a shadow GitHub Issue.

+JQL and issue fields are the real objects your process already uses
+Sprint and epic context keeps a quick-fix PR tied to the work item security cares about
+Pairs with **GitHub MCP Server** when code is on GitHub and tickets stay in Jira
Jira schemas vary by org. An agent will fill custom fields badly until you prompt for the required ones
Pick **Jira MCP Server** or **Linear MCP**, not both, unless you truly run two boards
7
Semgrep MCP logo

Semgrep MCP

4.6G2(55)

Teams that already run Semgrep in CI and want those findings inside the agent session.

Semgrep MCP screenshot
+The agent can query Semgrep findings on a path instead of ignoring CI annotations
+Useful in review: explain a finding, propose a fix, open the PR via **GitHub MCP Server**
+Rules you already own stay the source of truth. The MCP does not invent a second scanner
Untriaged rule packs produce more findings than a junior agent should auto-fix
It does not replace a human on high-severity security issues

Repos whose required checks run on CircleCI rather than GitHub Actions alone.

CircleCI MCP screenshot
+Failed job logs in the same session as the PR the agent is writing
+Stops the local-only loop when the only signal is a red badge
+Freemium CircleCI usage plus a free-to-run MCP is enough for a small team to try the loop
If you only use GitHub Actions, you may not need this server on day one
Log volume can drown the context window; ask for the failed step, not the whole job
Good value

CircleCI's pricing is fair for teams already invested in its ecosystem, with a generous Free tier offering 6,000 build minutes and 30x concurrency at $0/mo.

Watch out

Overage fees for extra build minutes

9
Figma MCP logo

Figma MCP

4.6G2(2,000)4.7Capterra(863)

UI work where the agent should see the file the designer named, not a screenshot in the ticket.

+Pulls structure and tokens from the Figma file instead of guessing spacing from a PNG
+Dev Mode-oriented workflows are the honest use: implement the frame, then PR via **GitHub MCP Server**
+Reduces spacing arguments during review because the frame is in the session
It will not clone a repo or merge a pull request. Do not install it as a GitHub substitute
A file with no Dev Mode hygiene gives the agent messy layers and a confident wrong layout
Good value

Figma's MCP pricing is generally fair, offering a robust free tier and tiered pricing that scales with organizational needs.

Watch out

AI credits are capped monthly, overages not specified

Agents that need a current example, a docs page, or a public snippet the repo does not contain.

+Neural and keyword search over the live web, including code-shaped queries
+Fills the gap **Context7 MCP** does not: pages that are not a library doc set
+Useful when you are evaluating an approach before you write the PR
It is search, not your git host. Ship still goes through **GitHub MCP Server** or **GitLab MCP Server**
Public snippets can be wrong or outdated; treat hits as leads, not as the patch
Good value

Exa MCP's pricing is fair and usage-based, with a generous free tier giving $20 sign-up credits and $10 monthly credits.

Watch out

Free credits expire if unused monthly

What an MCP server for GitHub and engineering actually is

An MCP server exposes a third-party API to an AI agent as typed tools. The official GitHub MCP Server turns "list my review requests," "create a pull request," and "comment on this issue" into calls the model can make with your token, instead of pasting git output into a chat window.

The rest of this list is not "more GitHub." It is the systems that surround a pull request. Merge requests live on GitLab MCP Server for GitLab shops. Production errors live in Sentry MCP. Work items live in Linear MCP or Jira MCP Server. Static findings live in Semgrep MCP. Build status lives in CircleCI MCP. Fresh framework docs live in Context7 MCP. Design context lives in Figma MCP. Search that is not your repo lives in Exa MCP.

That split is why this guide is not a clone of the devops MCP list (AWS, Kubernetes, Terraform) or the project-management MCP list (tickets in isolation). The unit of work here is the change: branch, review, check, ticket, error, and ship.

Why GitHub MCP without the adjacent servers stalls

An agent with only GitHub MCP Server can open a PR and still be blind. It cannot see the Sentry issue that caused the ticket. It cannot see the Linear cycle the PR is supposed to close. It cannot see the CircleCI job that failed on a flake. It will invent a library API because Context7 MCP is not in the session. Engineers then spend the review teaching the model facts that another server already had.

The other failure mode is wiring twenty MCP servers into one client and watching the model pick the wrong tool. Ten is already a lot. Start with GitHub MCP Server plus the one system of record your team actually uses for tickets (Linear MCP or Jira MCP Server) and the one system of record for errors (Sentry MCP). Add Semgrep MCP and CircleCI MCP when you trust the agent to read findings and pipeline output. Add Figma MCP only when the work is UI. Add Exa MCP when the repo is not enough.

Of 856 published tools in AI Agents, 56% are free or freemium, and 373 published tools on Toolradar are MCP servers or MCP-named. Plenty of options exist. The constraint is attention, not inventory.

Key Features to Look For

PR, issue, and repo tools on the git hostEssential

**GitHub MCP Server** or **GitLab MCP Server** must be able to read and write the objects you already live in: pull requests, merge requests, issues, comments, checks.

Token scope you would hand a junior engineerEssential

A fine-grained GitHub token or a GitLab token with least privilege is the real guardrail. An admin PAT in an agent config is an incident.

Error and ticket context next to the diffEssential

**Sentry MCP** plus **Linear MCP** or **Jira MCP Server** is how the agent stops guessing why the PR exists.

Checks and static analysis the agent can query

**CircleCI MCP** and **Semgrep MCP** turn a red CI badge and a SAST finding into tools, not screenshots pasted into chat.

Current library docs, not training-cutoff guesses

**Context7 MCP** exists because models invent APIs. Ground the agent in the docs for the version you actually installed.

Handoff and search without pretending they are git

**Figma MCP** is design context. **Exa MCP** is search. Neither replaces **GitHub MCP Server**.

Pricing Overview

MCP servers on this list

GitHub, GitLab, Sentry, Linear, Context7, Jira, Semgrep, CircleCI, Figma, Exa. Run locally or in the agent host.

$0 for the server
Underlying SaaS

A GitHub Free vs Team seat, a Sentry event quota, a Linear workspace. The MCP does not replace that invoice.

Your existing plan

Pricing Comparison

ToolMCP costYou already payJob in the loop
GitHub MCP Server$0GitHub plan + Actions minutesRepos, PRs, issues
GitLab MCP Server$0GitLab.com or self-managedMRs, issues, pipelines
Sentry MCP$0Sentry event quotaErrors into the agent
Linear MCP$0Linear workspaceIssues from the agent
Context7 MCP$0None requiredUp-to-date library docs
Jira MCP Server$0Jira / Atlassian seatTickets if you are not on Linear
Semgrep MCP$0Semgrep scan planSAST in the loop
CircleCI MCP$0CircleCI creditsPipeline status
Figma MCP$0Figma seatDesign context
Exa MCP$0Exa search usageWeb / code search

373 MCP-named tools in the catalog (2026-08-26 pull). This ten is the GitHub-centric engineering loop, not a dump of all 373. The MCP binary is free; a leaked classic PAT is the real cost.

Mistakes to Avoid

  • ×

    Skipping GitHub MCP Server and trying to drive PRs through a generic filesystem server

  • ×

    Dropping an admin GitHub PAT into the agent config instead of a fine-grained token on the two repos that matter

  • ×

    Installing Linear MCP and Jira MCP Server together because both appeared on a list

  • ×

    Treating Figma MCP as part of the git loop. It is handoff. Review still happens on the PR

  • ×

    Inventing a catalog size instead of using the 373 MCP-named published tools Toolradar actually counts

Expert Tips

  • Install GitHub MCP Server first, in read-heavy mode if your client supports it. Watch it list PRs and issues for a week before you let it create them.

  • The smallest useful bundle after GitHub is Sentry MCP plus one ticket server (Linear MCP or Jira MCP Server). That is the incident-to-PR loop.

  • Add Context7 MCP the first time you catch the model using a removed API. You will catch it.

  • Point Semgrep MCP at the same rules CI already runs. A second rule set in the agent session creates findings nobody else sees.

  • Keep Figma MCP off default agent profiles that do backend work. Load it for UI tickets only.

The Bottom Line

GitHub MCP Server is the engineering MCP. Add GitLab MCP Server only if GitLab is the host. Then attach the loop around the PR: Sentry MCP for errors, Linear MCP or Jira MCP Server for tickets, Semgrep MCP for findings, CircleCI MCP for pipelines, Context7 MCP for docs. Figma MCP is design handoff. Exa MCP is search. Toolradar tracks 373 published tools that are MCP servers or MCP-named, and the AI Agents category has 856 published tools with 56% free or freemium. The twelve other thematic MCP guides cover everything except this git and review loop. Start here, not with twenty servers.

Frequently Asked Questions

What is the best MCP server for GitHub?

GitHub MCP Server is the official one and the correct first install. It covers repositories, issues, pull requests, comments, and Actions-related objects. Scope it with a fine-grained token. Then add Sentry MCP, Linear MCP or Jira MCP Server, and Context7 MCP as the next three, depending on where errors, tickets, and docs live.

Do I need GitLab MCP if I already have GitHub MCP?

Only if the team actually hosts code on GitLab. GitLab MCP Server is for merge requests and issues on GitLab, including self-managed. Running both against the same feature work splits the agent's attention. Pick the host that owns the default branch.

How many MCP servers does Toolradar track?

Toolradar tracks 373 published tools that are MCP servers or MCP-named in the name, slug, tagline, or description. The AI Agents category, which is where these listings live, has 856 published tools, 56% of them free or freemium. The twelve existing thematic MCP guides cover databases, analytics, devops, search, comms, finance, sales, productivity, project management, ecommerce, support, and file storage. This guide is the git, PR, and issue loop.

Is Figma MCP a GitHub tool?

No. Figma MCP is ninth on this list because designers hand off frames the same week engineers open PRs, not because Figma is a git host. Use it to read the file, then implement and review through GitHub MCP Server.

Are these MCP servers free?

GitHub MCP Server, GitLab MCP Server, Sentry MCP, Linear MCP, Jira MCP Server, and Semgrep MCP are free servers on top of products you likely already pay for (or already have a free tier on). Context7 MCP, CircleCI MCP, Figma MCP, and Exa MCP are freemium at the product layer. The MCP process itself is not the line item. Tokens, search credits, and CI minutes are.

Related Guides

From the team behind Toolradar

Editorial content for AI startups

We turn AI product expertise into content that ranks, gets cited by LLMs, and reaches 720K+ tech buyers.

See how we work